Verified 1w ago Mainly llama.cpp 200 approved runs
Submissions
1,149
200 approved public runs
Models
28
unique models
Avg output
286tok/s
Best 3,216 tok/s
Avg prefill
333tok/s
Best 3,953 tok/s
Avg total
587tok/s
Best 6,953 tok/s
Avg TTFT
5,470ms
Best 77.2 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.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 8 requests sent at once (lmx --concurrency 8); 25.1 tok/s is the aggregate decode rate across all 8 streams, about 3.1 tok/s per stream, and the 4889 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 28 vCPU of an AMD EPYC 7343 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 77 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; 47380 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q4_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: 33.6, 33.1, 25.1 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. 4 requests sent at once (lmx --concurrency 4); 72.5 tok/s is the aggregate decode rate across all 4 streams, about 18.1 tok/s per stream, and the 2313 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 28 vCPU of an AMD EPYC 7343 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 77 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; 47380 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q4_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: 19.4, 72.5, 21.4 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); 51.9 tok/s is the aggregate decode rate across all 2 streams, about 25.9 tok/s per stream, and the 1593 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 28 vCPU of an AMD EPYC 7343 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 77 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; 47380 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q4_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: 51.9, 11.7, 11.3 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.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 28 vCPU of an AMD EPYC 7343 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 77 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; 47380 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q4_K_XL-00001-of-00003.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 36.8, 37.4, 38.8 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. 8 requests sent at once (batch 8, lmx --concurrency 8); 111.2 tok/s is the aggregate decode rate across all 8 streams, about 13.9 tok/s per stream, and the 3302 ms time to first token is under that load. Setup: KASA cuda-vm, 1 x NVIDIA RTX A6000 48 GB passed through to a VM (28 vCPU of an AMD EPYC 7343, 128 GiB), one card; vLLM 0.32.0 (vllm/vllm-openai@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34); max_num_seqs 32, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (316 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 8: 111.2, 111.4, 111.1 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config vq72-awq-eager, kasa-testing campaign ai/vm/cuda-vm/2026-10-26-a6000-localmaxxing. Selection declared before measuring: the median of the 1 measured requests whose lmx promptTokens is a whole number; the others report a fractional mean, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 8 requests sent at once (lmx --concurrency 8); 67 tok/s is the aggregate decode rate across all 8 streams, about 8.4 tok/s per stream, and the 2972 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 28 vCPU of an AMD EPYC 7343 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 56 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; 47424 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q3_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: 69.4, 67, 62.7 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); 44.4 tok/s is the aggregate decode rate across all 4 streams, about 11.1 tok/s per stream, and the 1615 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 28 vCPU of an AMD EPYC 7343 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 56 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; 47424 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q3_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: 49.3, 44.4, 42.7 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); 54.8 tok/s is the aggregate decode rate across all 2 streams, about 27.4 tok/s per stream, and the 1866 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 28 vCPU of an AMD EPYC 7343 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 111 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; 47390 MiB of VRAM in use after load, 16 generation threads, 28 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-Q4_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: 18.2, 16, 54.8 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 28 vCPU of an AMD EPYC 7343 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 111 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; 47390 MiB of VRAM in use after load, 16 generation threads, 28 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-Q4_K_XL-00001-of-00004.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 38.9, 38.8, 40.8 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. 2 requests sent at once (lmx --concurrency 2); 36.3 tok/s is the aggregate decode rate across all 2 streams, about 18.1 tok/s per stream, and the 1436 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 28 vCPU of an AMD EPYC 7343 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 56 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; 47424 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q3_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: 66.5, 35, 36.3 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 cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 28 vCPU of an AMD EPYC 7343 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 56 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; 47424 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q3_K_XL-00001-of-00003.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 45.1, 44.2, 43.6 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); 77.1 tok/s is the aggregate decode rate across all 8 streams, about 9.6 tok/s per stream, and the 3236 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 28 vCPU of an AMD EPYC 7343 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 81 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; 47228 MiB of VRAM in use after load, 16 generation threads, 28 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-IQ3_XXS-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: 132.8, 77.1, 55.9 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); 44.3 tok/s is the aggregate decode rate across all 2 streams, about 22.1 tok/s per stream, and the 906 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 28 vCPU of an AMD EPYC 7343 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 81 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; 47228 MiB of VRAM in use after load, 16 generation threads, 28 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-IQ3_XXS-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: 45.2, 80.8, 44.3 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.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 28 vCPU of an AMD EPYC 7343 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 81 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; 47228 MiB of VRAM in use after load, 16 generation threads, 28 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-IQ3_XXS-00001-of-00003.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 55.4, 55.1, 54.8 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. 8 requests sent at once (lmx --concurrency 8); 181.3 tok/s is the aggregate decode rate across all 8 streams, about 22.7 tok/s per stream, and the 1987 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 28 vCPU of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), all layers on the GPU (-ngl 99), 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q2_K_XL.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: 185.4, 181.3, 183.3 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.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 4 requests sent at once (lmx --concurrency 4); 157 tok/s is the aggregate decode rate across all 4 streams, about 39.2 tok/s per stream, and the 975 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 28 vCPU of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), all layers on the GPU (-ngl 99), 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q2_K_XL.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: 158.1, 157, 157.6 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 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); 120 tok/s is the aggregate decode rate across all 2 streams, about 60.0 tok/s per stream, and the 669 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 28 vCPU of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), all layers on the GPU (-ngl 99), 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q2_K_XL.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: 121.9, 120.5, 120 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 28 vCPU of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), all layers on the GPU (-ngl 99), 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q2_K_XL.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 73, 72.7, 72 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); 15.9 tok/s is the aggregate decode rate across all 8 streams, about 2.0 tok/s per stream, and the 5314 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 28 vCPU of an AMD EPYC 7343 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 113 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, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF ggml-org/GLM-4.6V-GGUF @ 7de678a269f576ae4090efbf59b558a993386fff, file GLM-4.6V-Q8_0-00001-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: 15.1, 15.9, 46.7 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 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. 4 requests sent at once (batch 4, lmx --concurrency 4); 59.5 tok/s is the aggregate decode rate across all 4 streams, about 14.9 tok/s per stream, and the 1689 ms time to first token is under that load. Setup: KASA cuda-vm, 1 x NVIDIA RTX A6000 48 GB passed through to a VM (28 vCPU of an AMD EPYC 7343, 128 GiB), one card; vLLM 0.32.0 (vllm/vllm-openai@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34); max_num_seqs 32, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (317 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 4: 59.5, 59.5, 59.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config vq72-awq-eager, kasa-testing campaign ai/vm/cuda-vm/2026-10-26-a6000-localmaxxing. Selection declared before measuring: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional mean, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 32 requests sent at once (batch 32, lmx --concurrency 32); 158.6 tok/s is the aggregate decode rate across all 32 streams, about 5.0 tok/s per stream, and the 9444 ms time to first token is under that load. Setup: KASA cuda-vm, 1 x NVIDIA RTX A6000 48 GB passed through to a VM (28 vCPU of an AMD EPYC 7343, 128 GiB), one card; vLLM 0.32.0 (vllm/vllm-openai@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34); max_num_seqs 32, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (317 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 32: 158.3, 158.6, 159 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config vq72-awq-eager, kasa-testing campaign ai/vm/cuda-vm/2026-10-26-a6000-localmaxxing. Selection declared before measuring: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional mean, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 2 requests sent at once (batch 2, lmx --concurrency 2); 31.2 tok/s is the aggregate decode rate across all 2 streams, about 15.6 tok/s per stream, and the 676 ms time to first token is under that load. Setup: KASA cuda-vm, 1 x NVIDIA RTX A6000 48 GB passed through to a VM (28 vCPU of an AMD EPYC 7343, 128 GiB), one card; vLLM 0.32.0 (vllm/vllm-openai@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34); max_num_seqs 32, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (317 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 2: 31.5, 31.2, 31.5 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config vq72-awq-eager, kasa-testing campaign ai/vm/cuda-vm/2026-10-26-a6000-localmaxxing. Selection declared before measuring: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional mean, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 16 requests sent at once (batch 16, lmx --concurrency 16); 178 tok/s is the aggregate decode rate across all 16 streams, about 11.1 tok/s per stream, and the 4960 ms time to first token is under that load. Setup: KASA cuda-vm, 1 x NVIDIA RTX A6000 48 GB passed through to a VM (28 vCPU of an AMD EPYC 7343, 128 GiB), one card; vLLM 0.32.0 (vllm/vllm-openai@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34); max_num_seqs 32, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (317 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 16: 178, 177.5, 178.1 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config vq72-awq-eager, kasa-testing campaign ai/vm/cuda-vm/2026-10-26-a6000-localmaxxing.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 4 requests sent at once (lmx --concurrency 4); 5.7 tok/s is the aggregate decode rate across all 4 streams, about 1.4 tok/s per stream, and the 3979 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 28 vCPU of an AMD EPYC 7343 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 113 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, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF ggml-org/GLM-4.6V-GGUF @ 7de678a269f576ae4090efbf59b558a993386fff, file GLM-4.6V-Q8_0-00001-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: 9.4, 5.7, 12.3 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.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 28 vCPU of an AMD EPYC 7343 and 128 GiB. Config vq72-awq-eager, vLLM 0.32.0. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 16.3, 16.1, 16.3 tok/s; this is the median (declared rule, tie to the earlier). Prompt/output samples, engineTimingsRaw and MTP counts were added from the lmx run file and exact per-request vLLM /metrics deltas, because lmx does not send them.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 8 requests sent at once (lmx --concurrency 8); 25.1 tok/s is the aggregate decode rate across all 8 streams, about 3.1 tok/s per stream, and the 4889 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 28 vCPU of an AMD EPYC 7343 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 77 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; 47380 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q4_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: 33.6, 33.1, 25.1 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. 4 requests sent at once (lmx --concurrency 4); 72.5 tok/s is the aggregate decode rate across all 4 streams, about 18.1 tok/s per stream, and the 2313 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 28 vCPU of an AMD EPYC 7343 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 77 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; 47380 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q4_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: 19.4, 72.5, 21.4 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); 51.9 tok/s is the aggregate decode rate across all 2 streams, about 25.9 tok/s per stream, and the 1593 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 28 vCPU of an AMD EPYC 7343 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 77 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; 47380 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q4_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: 51.9, 11.7, 11.3 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.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 28 vCPU of an AMD EPYC 7343 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 77 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; 47380 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q4_K_XL-00001-of-00003.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 36.8, 37.4, 38.8 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. 8 requests sent at once (batch 8, lmx --concurrency 8); 111.2 tok/s is the aggregate decode rate across all 8 streams, about 13.9 tok/s per stream, and the 3302 ms time to first token is under that load. Setup: KASA cuda-vm, 1 x NVIDIA RTX A6000 48 GB passed through to a VM (28 vCPU of an AMD EPYC 7343, 128 GiB), one card; vLLM 0.32.0 (vllm/vllm-openai@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34); max_num_seqs 32, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (316 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 8: 111.2, 111.4, 111.1 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config vq72-awq-eager, kasa-testing campaign ai/vm/cuda-vm/2026-10-26-a6000-localmaxxing. Selection declared before measuring: the median of the 1 measured requests whose lmx promptTokens is a whole number; the others report a fractional mean, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 8 requests sent at once (lmx --concurrency 8); 67 tok/s is the aggregate decode rate across all 8 streams, about 8.4 tok/s per stream, and the 2972 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 28 vCPU of an AMD EPYC 7343 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 56 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; 47424 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q3_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: 69.4, 67, 62.7 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); 44.4 tok/s is the aggregate decode rate across all 4 streams, about 11.1 tok/s per stream, and the 1615 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 28 vCPU of an AMD EPYC 7343 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 56 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; 47424 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q3_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: 49.3, 44.4, 42.7 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); 54.8 tok/s is the aggregate decode rate across all 2 streams, about 27.4 tok/s per stream, and the 1866 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 28 vCPU of an AMD EPYC 7343 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 111 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; 47390 MiB of VRAM in use after load, 16 generation threads, 28 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-Q4_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: 18.2, 16, 54.8 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 28 vCPU of an AMD EPYC 7343 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 111 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; 47390 MiB of VRAM in use after load, 16 generation threads, 28 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-Q4_K_XL-00001-of-00004.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 38.9, 38.8, 40.8 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. 2 requests sent at once (lmx --concurrency 2); 36.3 tok/s is the aggregate decode rate across all 2 streams, about 18.1 tok/s per stream, and the 1436 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 28 vCPU of an AMD EPYC 7343 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 56 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; 47424 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q3_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: 66.5, 35, 36.3 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 cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 28 vCPU of an AMD EPYC 7343 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 56 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; 47424 MiB of VRAM in use after load, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q3_K_XL-00001-of-00003.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 45.1, 44.2, 43.6 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); 77.1 tok/s is the aggregate decode rate across all 8 streams, about 9.6 tok/s per stream, and the 3236 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 28 vCPU of an AMD EPYC 7343 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 81 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; 47228 MiB of VRAM in use after load, 16 generation threads, 28 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-IQ3_XXS-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: 132.8, 77.1, 55.9 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); 44.3 tok/s is the aggregate decode rate across all 2 streams, about 22.1 tok/s per stream, and the 906 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 28 vCPU of an AMD EPYC 7343 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 81 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; 47228 MiB of VRAM in use after load, 16 generation threads, 28 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-IQ3_XXS-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: 45.2, 80.8, 44.3 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.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 28 vCPU of an AMD EPYC 7343 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 81 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; 47228 MiB of VRAM in use after load, 16 generation threads, 28 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-IQ3_XXS-00001-of-00003.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 55.4, 55.1, 54.8 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. 8 requests sent at once (lmx --concurrency 8); 181.3 tok/s is the aggregate decode rate across all 8 streams, about 22.7 tok/s per stream, and the 1987 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 28 vCPU of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), all layers on the GPU (-ngl 99), 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q2_K_XL.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: 185.4, 181.3, 183.3 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.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 4 requests sent at once (lmx --concurrency 4); 157 tok/s is the aggregate decode rate across all 4 streams, about 39.2 tok/s per stream, and the 975 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 28 vCPU of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), all layers on the GPU (-ngl 99), 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q2_K_XL.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: 158.1, 157, 157.6 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 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); 120 tok/s is the aggregate decode rate across all 2 streams, about 60.0 tok/s per stream, and the 669 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 28 vCPU of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), all layers on the GPU (-ngl 99), 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q2_K_XL.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: 121.9, 120.5, 120 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 28 vCPU of an AMD EPYC 7343 and 128 GiB. llama.cpp llama-server b11889 (CUDA image ghcr.io/ggml-org/llama.cpp), all layers on the GPU (-ngl 99), 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF unsloth/Qwen3.5-122B-A10B-GGUF @ 51eab4d59d53f573fb9206cb3ce613f1d0aa392b, file Qwen3.5-122B-A10B-UD-Q2_K_XL.gguf. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 73, 72.7, 72 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); 15.9 tok/s is the aggregate decode rate across all 8 streams, about 2.0 tok/s per stream, and the 5314 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 28 vCPU of an AMD EPYC 7343 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 113 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, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF ggml-org/GLM-4.6V-GGUF @ 7de678a269f576ae4090efbf59b558a993386fff, file GLM-4.6V-Q8_0-00001-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: 15.1, 15.9, 46.7 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 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. 4 requests sent at once (batch 4, lmx --concurrency 4); 59.5 tok/s is the aggregate decode rate across all 4 streams, about 14.9 tok/s per stream, and the 1689 ms time to first token is under that load. Setup: KASA cuda-vm, 1 x NVIDIA RTX A6000 48 GB passed through to a VM (28 vCPU of an AMD EPYC 7343, 128 GiB), one card; vLLM 0.32.0 (vllm/vllm-openai@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34); max_num_seqs 32, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (317 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 4: 59.5, 59.5, 59.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config vq72-awq-eager, kasa-testing campaign ai/vm/cuda-vm/2026-10-26-a6000-localmaxxing. Selection declared before measuring: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional mean, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 32 requests sent at once (batch 32, lmx --concurrency 32); 158.6 tok/s is the aggregate decode rate across all 32 streams, about 5.0 tok/s per stream, and the 9444 ms time to first token is under that load. Setup: KASA cuda-vm, 1 x NVIDIA RTX A6000 48 GB passed through to a VM (28 vCPU of an AMD EPYC 7343, 128 GiB), one card; vLLM 0.32.0 (vllm/vllm-openai@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34); max_num_seqs 32, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (317 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 32: 158.3, 158.6, 159 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config vq72-awq-eager, kasa-testing campaign ai/vm/cuda-vm/2026-10-26-a6000-localmaxxing. Selection declared before measuring: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional mean, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 2 requests sent at once (batch 2, lmx --concurrency 2); 31.2 tok/s is the aggregate decode rate across all 2 streams, about 15.6 tok/s per stream, and the 676 ms time to first token is under that load. Setup: KASA cuda-vm, 1 x NVIDIA RTX A6000 48 GB passed through to a VM (28 vCPU of an AMD EPYC 7343, 128 GiB), one card; vLLM 0.32.0 (vllm/vllm-openai@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34); max_num_seqs 32, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (317 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 2: 31.5, 31.2, 31.5 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config vq72-awq-eager, kasa-testing campaign ai/vm/cuda-vm/2026-10-26-a6000-localmaxxing. Selection declared before measuring: the median of the 2 measured requests whose lmx promptTokens is a whole number; the others report a fractional mean, which LocalMaxxing refuses.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 16 requests sent at once (batch 16, lmx --concurrency 16); 178 tok/s is the aggregate decode rate across all 16 streams, about 11.1 tok/s per stream, and the 4960 ms time to first token is under that load. Setup: KASA cuda-vm, 1 x NVIDIA RTX A6000 48 GB passed through to a VM (28 vCPU of an AMD EPYC 7343, 128 GiB), one card; vLLM 0.32.0 (vllm/vllm-openai@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34); max_num_seqs 32, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (317 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 16: 178, 177.5, 178.1 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config vq72-awq-eager, kasa-testing campaign ai/vm/cuda-vm/2026-10-26-a6000-localmaxxing.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 4 requests sent at once (lmx --concurrency 4); 5.7 tok/s is the aggregate decode rate across all 4 streams, about 1.4 tok/s per stream, and the 3979 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 28 vCPU of an AMD EPYC 7343 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 113 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, 16 generation threads, 28 batch threads, flash attention on, 8 server slots of 4096 tokens each, context 32768. GGUF ggml-org/GLM-4.6V-GGUF @ 7de678a269f576ae4090efbf59b558a993386fff, file GLM-4.6V-Q8_0-00001-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: 9.4, 5.7, 12.3 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.
KASA cuda-vm: 1 x NVIDIA RTX A6000 48 GB (PCIe passthrough, driver 615.71.10) in a VM with 28 vCPU of an AMD EPYC 7343 and 128 GiB. Config vq72-awq-eager, vLLM 0.32.0. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 16.3, 16.1, 16.3 tok/s; this is the median (declared rule, tie to the earlier). Prompt/output samples, engineTimingsRaw and MTP counts were added from the lmx run file and exact per-request vLLM /metrics deltas, because lmx does not send them.
