Verified 3w ago Mainly llama.cpp 200 approved runs
Submissions
2,318
200 approved public runs
Models
16
unique models
Avg output
24.0tok/s
Best 103 tok/s
Avg prefill
72.7tok/s
Best 281 tok/s
Avg total
43.4tok/s
Best 215 tok/s
Avg TTFT
10,220ms
Best 1,110 ms
Hardware
2
distinct setups
Matrix runs
36
speed test matrices
Matrix runs36+
Matrix runs 36
A matrix run measures prompt reading (prefill) and generation (decode) speed in one go, across several context lengths and numbers of simultaneous requests. Showing the newest 10.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 108 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47800 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768, KV cache q8_0 keys / q8_0 values (-ctk/-ctv). GGUF unsloth/GLM-5.3-Flash-GGUF @ a38483c8cd5df544f53d70fb281afe97369d5ab6, file GLM-5.3-Flash-UD-Q2_K_XL-00001-of-00004.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 7.8, 8, 7.8 tok/s; this is the median (tie to the earlier). Prompt and output samples were added from the lmx run file, engineTimingsRaw from exact llama-server /metrics deltas around the request.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 3 requests sent at once (lmx --concurrency 3); 7.1 tok/s is the aggregate decode rate across all 3 streams, about 2.4 tok/s per stream, and the 8315 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 16 server slots of 4096 tokens each, context 65536. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 5.9, 7.6, 7.1 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47628 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768, KV cache q8_0 keys / q8_0 values (-ctk/-ctv). GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 13, 14.4, 13.1 tok/s; this is the median (tie to the earlier). Prompt and output samples were added from the lmx run file, engineTimingsRaw from exact llama-server /metrics deltas around the request.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 4 requests sent at once (lmx --concurrency 4); 9.6 tok/s is the aggregate decode rate across all 4 streams, about 2.4 tok/s per stream, and the 8304 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 126 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47220 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768, KV cache q8_0 keys / q8_0 values (-ctk/-ctv). GGUF ggml-org/MiMo-V2.6-Flash-RL-GGUF @ f1ef9bfdc3706af3f807752765e52b885e113aae, file MiMo-V2.6-Flash-RL-Q2_K-00001-of-00002.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 17.4, 17.6, 9.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 126 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47220 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768, KV cache q8_0 keys / q8_0 values (-ctk/-ctv). GGUF ggml-org/MiMo-V2.6-Flash-RL-GGUF @ f1ef9bfdc3706af3f807752765e52b885e113aae, file MiMo-V2.6-Flash-RL-Q2_K-00001-of-00002.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 17.4, 18.2, 18.9 tok/s; this is the median (tie to the earlier). Prompt and output samples were added from the lmx run file, engineTimingsRaw from exact llama-server /metrics deltas around the request.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 2 requests sent at once (lmx --concurrency 2); 7.3 tok/s is the aggregate decode rate across all 2 streams, about 3.6 tok/s per stream, and the 6096 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 16 server slots of 4096 tokens each, context 65536. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 7, 7.3, 9.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 16 server slots of 4096 tokens each, context 65536. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 7.2, 7.3, 7.2 tok/s; this is the median (tie to the earlier). Prompt and output samples were added from the lmx run file, engineTimingsRaw from exact llama-server /metrics deltas around the request.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 8 requests sent at once (lmx --concurrency 8); 10.2 tok/s is the aggregate decode rate across all 8 streams, about 1.3 tok/s per stream, and the 30963 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 10.2, 10.3, 10.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 7 requests sent at once (lmx --concurrency 7); 9.8 tok/s is the aggregate decode rate across all 7 streams, about 1.4 tok/s per stream, and the 29675 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 9.8, 7.5, 10.1 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 2 requests sent at once (lmx --concurrency 2); 1.3 tok/s is the aggregate decode rate across all 2 streams, about 0.7 tok/s per stream, and the 6700 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 120 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47870 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/GLM-5.3-Flash-GGUF @ a38483c8cd5df544f53d70fb281afe97369d5ab6, file GLM-5.3-Flash-UD-IQ3_XXS-00001-of-00004.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 1.3, 2.8, 3.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 6 requests sent at once (lmx --concurrency 6); 7.9 tok/s is the aggregate decode rate across all 6 streams, about 1.3 tok/s per stream, and the 17369 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 10.1, 8.2, 7.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 1 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 2 requests sent at once (lmx --concurrency 2); 1.9 tok/s is the aggregate decode rate across all 2 streams, about 0.9 tok/s per stream, and the 10051 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 108 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47414 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/GLM-5.3-Flash-GGUF @ a38483c8cd5df544f53d70fb281afe97369d5ab6, file GLM-5.3-Flash-UD-Q2_K_XL-00001-of-00004.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 2.3, 7.4, 1.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 1 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 120 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47870 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/GLM-5.3-Flash-GGUF @ a38483c8cd5df544f53d70fb281afe97369d5ab6, file GLM-5.3-Flash-UD-IQ3_XXS-00001-of-00004.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 6.5, 7.3, 7 tok/s; this is the median (tie to the earlier). Prompt and output samples were added from the lmx run file, engineTimingsRaw from exact llama-server /metrics deltas around the request.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 5 requests sent at once (lmx --concurrency 5); 7.2 tok/s is the aggregate decode rate across all 5 streams, about 1.4 tok/s per stream, and the 14652 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 7.3, 6.4, 7.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 4 requests sent at once (lmx --concurrency 4); 8.3 tok/s is the aggregate decode rate across all 4 streams, about 2.1 tok/s per stream, and the 11752 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 6.7, 8.3, 8.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 1 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 6 requests sent at once (lmx --concurrency 6); 25.7 tok/s is the aggregate decode rate across all 6 streams, about 4.3 tok/s per stream, and the 7676 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 71 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47142 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF @ 99144d1d7a318f045f1412a85bbb807f3c8a17a0, file NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Q4_0.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 30.6, 25.7, 35.7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 3 requests sent at once (lmx --concurrency 3); 25.2 tok/s is the aggregate decode rate across all 3 streams, about 8.4 tok/s per stream, and the 3098 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 71 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47142 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF @ 99144d1d7a318f045f1412a85bbb807f3c8a17a0, file NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Q4_0.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 30.5, 19.5, 25.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 6 requests sent at once (lmx --concurrency 6); 103.3 tok/s is the aggregate decode rate across all 6 streams, about 17.2 tok/s per stream, and the 1948 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 78 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47388 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.8-Flash-Next-GGUF @ 766911a6b7369840a91dbcd95f9f997acaab6cd6, file Qwen3.8-Flash-Next-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 103.3, 95.4, 105.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 3 requests sent at once (lmx --concurrency 3); 75.9 tok/s is the aggregate decode rate across all 3 streams, about 25.3 tok/s per stream, and the 1237 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 78 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47388 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.8-Flash-Next-GGUF @ 766911a6b7369840a91dbcd95f9f997acaab6cd6, file Qwen3.8-Flash-Next-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 77.6, 75.9, 75.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 3 requests sent at once (lmx --concurrency 3); 7 tok/s is the aggregate decode rate across all 3 streams, about 2.3 tok/s per stream, and the 9930 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 5.9, 7.5, 7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 6 requests sent at once (lmx --concurrency 6); 16.4 tok/s is the aggregate decode rate across all 6 streams, about 2.7 tok/s per stream, and the 4726 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 90 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47376 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-IQ2_M-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 16.4, 13.1, 14.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 1 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 3 requests sent at once (lmx --concurrency 3); 10.7 tok/s is the aggregate decode rate across all 3 streams, about 3.6 tok/s per stream, and the 5002 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 90 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47376 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-IQ2_M-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 10.1, 10.7, 12.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 2 requests sent at once (lmx --concurrency 2); 7.1 tok/s is the aggregate decode rate across all 2 streams, about 3.5 tok/s per stream, and the 6080 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 8, 7.1, 6.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 1 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 6 requests sent at once (lmx --concurrency 6); 36 tok/s is the aggregate decode rate across all 6 streams, about 6.0 tok/s per stream, and the 4297 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 118 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47748 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF ggml-org/Step-3.5-Flash-GGUF @ c439ad64d667632f3e1507757cbe225b6a46e6b8, file Step-3.5-Flash-Q4_K.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 36, 29, 37.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 3 requests sent at once (lmx --concurrency 3); 23.2 tok/s is the aggregate decode rate across all 3 streams, about 7.7 tok/s per stream, and the 3689 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 118 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47748 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF ggml-org/Step-3.5-Flash-GGUF @ c439ad64d667632f3e1507757cbe225b6a46e6b8, file Step-3.5-Flash-Q4_K.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 24, 13.2, 23.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 108 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47800 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768, KV cache q8_0 keys / q8_0 values (-ctk/-ctv). GGUF unsloth/GLM-5.3-Flash-GGUF @ a38483c8cd5df544f53d70fb281afe97369d5ab6, file GLM-5.3-Flash-UD-Q2_K_XL-00001-of-00004.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 7.8, 8, 7.8 tok/s; this is the median (tie to the earlier). Prompt and output samples were added from the lmx run file, engineTimingsRaw from exact llama-server /metrics deltas around the request.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 3 requests sent at once (lmx --concurrency 3); 7.1 tok/s is the aggregate decode rate across all 3 streams, about 2.4 tok/s per stream, and the 8315 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 16 server slots of 4096 tokens each, context 65536. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 5.9, 7.6, 7.1 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47628 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768, KV cache q8_0 keys / q8_0 values (-ctk/-ctv). GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 13, 14.4, 13.1 tok/s; this is the median (tie to the earlier). Prompt and output samples were added from the lmx run file, engineTimingsRaw from exact llama-server /metrics deltas around the request.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 4 requests sent at once (lmx --concurrency 4); 9.6 tok/s is the aggregate decode rate across all 4 streams, about 2.4 tok/s per stream, and the 8304 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 126 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47220 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768, KV cache q8_0 keys / q8_0 values (-ctk/-ctv). GGUF ggml-org/MiMo-V2.6-Flash-RL-GGUF @ f1ef9bfdc3706af3f807752765e52b885e113aae, file MiMo-V2.6-Flash-RL-Q2_K-00001-of-00002.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 17.4, 17.6, 9.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 126 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47220 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768, KV cache q8_0 keys / q8_0 values (-ctk/-ctv). GGUF ggml-org/MiMo-V2.6-Flash-RL-GGUF @ f1ef9bfdc3706af3f807752765e52b885e113aae, file MiMo-V2.6-Flash-RL-Q2_K-00001-of-00002.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 17.4, 18.2, 18.9 tok/s; this is the median (tie to the earlier). Prompt and output samples were added from the lmx run file, engineTimingsRaw from exact llama-server /metrics deltas around the request.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 2 requests sent at once (lmx --concurrency 2); 7.3 tok/s is the aggregate decode rate across all 2 streams, about 3.6 tok/s per stream, and the 6096 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 16 server slots of 4096 tokens each, context 65536. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 7, 7.3, 9.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 16 server slots of 4096 tokens each, context 65536. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 7.2, 7.3, 7.2 tok/s; this is the median (tie to the earlier). Prompt and output samples were added from the lmx run file, engineTimingsRaw from exact llama-server /metrics deltas around the request.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 8 requests sent at once (lmx --concurrency 8); 10.2 tok/s is the aggregate decode rate across all 8 streams, about 1.3 tok/s per stream, and the 30963 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 10.2, 10.3, 10.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 7 requests sent at once (lmx --concurrency 7); 9.8 tok/s is the aggregate decode rate across all 7 streams, about 1.4 tok/s per stream, and the 29675 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 9.8, 7.5, 10.1 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 2 requests sent at once (lmx --concurrency 2); 1.3 tok/s is the aggregate decode rate across all 2 streams, about 0.7 tok/s per stream, and the 6700 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 120 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47870 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/GLM-5.3-Flash-GGUF @ a38483c8cd5df544f53d70fb281afe97369d5ab6, file GLM-5.3-Flash-UD-IQ3_XXS-00001-of-00004.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 1.3, 2.8, 3.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 6 requests sent at once (lmx --concurrency 6); 7.9 tok/s is the aggregate decode rate across all 6 streams, about 1.3 tok/s per stream, and the 17369 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 10.1, 8.2, 7.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 1 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 2 requests sent at once (lmx --concurrency 2); 1.9 tok/s is the aggregate decode rate across all 2 streams, about 0.9 tok/s per stream, and the 10051 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 108 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47414 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/GLM-5.3-Flash-GGUF @ a38483c8cd5df544f53d70fb281afe97369d5ab6, file GLM-5.3-Flash-UD-Q2_K_XL-00001-of-00004.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 2.3, 7.4, 1.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 1 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 120 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47870 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/GLM-5.3-Flash-GGUF @ a38483c8cd5df544f53d70fb281afe97369d5ab6, file GLM-5.3-Flash-UD-IQ3_XXS-00001-of-00004.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 6.5, 7.3, 7 tok/s; this is the median (tie to the earlier). Prompt and output samples were added from the lmx run file, engineTimingsRaw from exact llama-server /metrics deltas around the request.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 5 requests sent at once (lmx --concurrency 5); 7.2 tok/s is the aggregate decode rate across all 5 streams, about 1.4 tok/s per stream, and the 14652 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 7.3, 6.4, 7.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 4 requests sent at once (lmx --concurrency 4); 8.3 tok/s is the aggregate decode rate across all 4 streams, about 2.1 tok/s per stream, and the 11752 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 6.7, 8.3, 8.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 1 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 6 requests sent at once (lmx --concurrency 6); 25.7 tok/s is the aggregate decode rate across all 6 streams, about 4.3 tok/s per stream, and the 7676 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 71 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47142 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF @ 99144d1d7a318f045f1412a85bbb807f3c8a17a0, file NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Q4_0.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 30.6, 25.7, 35.7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 3 requests sent at once (lmx --concurrency 3); 25.2 tok/s is the aggregate decode rate across all 3 streams, about 8.4 tok/s per stream, and the 3098 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 71 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47142 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF @ 99144d1d7a318f045f1412a85bbb807f3c8a17a0, file NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-Q4_0.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 30.5, 19.5, 25.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 6 requests sent at once (lmx --concurrency 6); 103.3 tok/s is the aggregate decode rate across all 6 streams, about 17.2 tok/s per stream, and the 1948 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 78 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47388 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.8-Flash-Next-GGUF @ 766911a6b7369840a91dbcd95f9f997acaab6cd6, file Qwen3.8-Flash-Next-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 103.3, 95.4, 105.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 3 requests sent at once (lmx --concurrency 3); 75.9 tok/s is the aggregate decode rate across all 3 streams, about 25.3 tok/s per stream, and the 1237 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 78 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47388 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.8-Flash-Next-GGUF @ 766911a6b7369840a91dbcd95f9f997acaab6cd6, file Qwen3.8-Flash-Next-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 77.6, 75.9, 75.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 3 requests sent at once (lmx --concurrency 3); 7 tok/s is the aggregate decode rate across all 3 streams, about 2.3 tok/s per stream, and the 9930 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 5.9, 7.5, 7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 6 requests sent at once (lmx --concurrency 6); 16.4 tok/s is the aggregate decode rate across all 6 streams, about 2.7 tok/s per stream, and the 4726 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 90 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47376 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-IQ2_M-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 16.4, 13.1, 14.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 1 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 3 requests sent at once (lmx --concurrency 3); 10.7 tok/s is the aggregate decode rate across all 3 streams, about 3.6 tok/s per stream, and the 5002 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 90 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47376 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-IQ2_M-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 10.1, 10.7, 12.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 2 requests sent at once (lmx --concurrency 2); 7.1 tok/s is the aggregate decode rate across all 2 streams, about 3.5 tok/s per stream, and the 6080 ms time to first token is under that load. KASA b65-vm: 1 x Intel Arc Pro B65 32 GB (PCIe passthrough, xe driver, Ubuntu 26.04 kernel 7.0) in a VM with 16 vCPU pinned to 8 cores (16 threads) of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11913 (image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 96 GB model does not fit the 32 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU; the exact split is not logged at the default verbosity, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/DeepSeek-V4-Flash-0731-GGUF @ fbbb5b93fb787c21338159b0af3318bb3f4d9768, file DeepSeek-V4-Flash-0731-UD-Q2_K_XL-00001-of-00003.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 8, 7.1, 6.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Selection rule applied to every run in this round: the median of the 1 measured requests whose lmx promptTokens is a whole number; the others report a fractional median, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 6 requests sent at once (lmx --concurrency 6); 36 tok/s is the aggregate decode rate across all 6 streams, about 6.0 tok/s per stream, and the 4297 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 118 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47748 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF ggml-org/Step-3.5-Flash-GGUF @ c439ad64d667632f3e1507757cbe225b6a46e6b8, file Step-3.5-Flash-Q4_K.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 36, 29, 37.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 3 requests sent at once (lmx --concurrency 3); 23.2 tok/s is the aggregate decode rate across all 3 streams, about 7.7 tok/s per stream, and the 3689 ms time to first token is under that load. KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 16 vCPU pinned to 8 of the 16 cores (16 threads) of an AMD EPYC 7343 (another VM holds the other 8 cores) and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), placement by llama.cpp's default --fit (1 GiB margin): the 118 GB model does not fit the 48 GB GPU, so llama.cpp's --fit kept the layers on the GPU and placed part of the MoE expert weights in system RAM on the CPU (loaded with -lm none, so the CPU part is in ordinary RAM, not memory-mapped); the exact split is not logged at the default verbosity; 47748 MiB of VRAM in use after load, 8 generation threads, 16 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF ggml-org/Step-3.5-Flash-GGUF @ c439ad64d667632f3e1507757cbe225b6a46e6b8, file Step-3.5-Flash-Q4_K.gguf. Each request: LocalMaxxing's canonical reasoning-v1 prompt with a unique nonce, 256 output tokens. One discarded warm-up batch, then 3 measured batches: 24, 13.2, 23.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1.
