Matrix runs
A matrix run measures prompt reading (prefill) and generation (decode) speed in one go, across several context lengths and numbers of simultaneous requests. Run one with lmx speed-test matrix run.
All runs 738
Newest first. Page 2 of 30.- Qwen3.5-122B-A10B-GPTQ-Int4Intel Arc Pro B65 32GB ×4 · vllm 0.31.0+xpu.llmscaler · GPTQ-4bit6 contexts × 5 concurrency levels
- Peak decode
- 539.6 tok/s
- Peak prefill
- 6,010 tok/s
- Muse-Glimmer-30BIntel Arc Pro B65 32GB ×4 · vllm 0.26.1.dev0+xpu.llmscaler-1.5 · FP84 contexts × 5 concurrency levels
- Peak decode
- 343.7 tok/s
- Peak prefill
- 2,679 tok/s
- Muse-Glimmer-30BIntel Arc Pro B65 32GB ×4 · vllm 0.31.0+xpu.llmscaler · FP84 contexts × 5 concurrency levels
- Peak decode
- 476.4 tok/s
- Peak prefill
- 2,709 tok/s
- Peak decode
- 399.3 tok/s
- Peak prefill
- 3,898 tok/s
- Gemma-4-31B-IT-NVFP4H200 NVL 141GB ×2 · sglang 0.5.6.post3.dev9335+gdd15fb57b · NVFP46 contexts × 8 concurrency levels
- Peak decode
- 170.5 tok/s
- Peak prefill
- 6,589 tok/s
- Peak decode
- 595.7 tok/s
- Peak prefill
- 9,041 tok/s
- Peak decode
- 1,670.0 tok/s
- Peak prefill
- 14.3K tok/s
- Qwen3.5-9BIntel Arc Pro B65 32GB ×4 · vllm 0.31.0+xpu.llmscaler · FP86 contexts × 5 concurrency levels
- Peak decode
- 1,568.3 tok/s
- Peak prefill
- 10.6K tok/s
- Peak decode
- 1,132.8 tok/s
- Peak prefill
- 10.4K tok/s
- DeepSeek-V4.1-Flash-1017H200 NVL 141GB ×2 · vllm 0.32.0 · MXFP4+FP86 contexts × 8 concurrency levels
- Peak decode
- 303.7 tok/s
- Peak prefill
- 2,844 tok/s
- Qwen3.8-27BIntel Arc Pro B65 32GB ×4 · vllm 0.31.0+xpu.llmscaler · INT46 contexts × 5 concurrency levels
- Peak decode
- 468.2 tok/s
- Peak prefill
- 3,748 tok/s
- Peak decode
- 404.2 tok/s
- Peak prefill
- 2,099 tok/s
- Peak decode
- 389.5 tok/s
- Peak prefill
- 3,699 tok/s
- Qwen3.6-35B-A3BIntel Arc Pro B65 32GB ×4 · vllm 0.31.0+xpu.llmscaler · INT46 contexts × 5 concurrency levels
- Peak decode
- 833.4 tok/s
- Peak prefill
- 12.3K tok/s
- Qwen3.6-35B-A3BIntel Arc Pro B65 32GB ×4 · vllm 0.31.0+xpu.llmscaler · FP86 contexts × 5 concurrency levels
- Peak decode
- 1,030.5 tok/s
- Peak prefill
- 12.8K tok/s
- Peak decode
- 855.6 tok/s
- Peak prefill
- 11.9K tok/s
- Qwen3.6-27BIntel Arc Pro B65 32GB ×4 · vllm 0.31.0+xpu.llmscaler · INT46 contexts × 5 concurrency levels
- Peak decode
- 471.7 tok/s
- Peak prefill
- 3,752 tok/s
- Qwen3.8-27B-W3Rot-INT3-Paiton-RDNA4Radeon AI Pro R9700 32GB · vllm 0.29.0 · W3A43 contexts × 2 concurrency levels
- Peak decode
- 543.7 tok/s
- Peak prefill
- 3,973 tok/s
- Qwen3.6-27BIntel Arc Pro B65 32GB ×4 · vllm 0.31.0+xpu.llmscaler · FP86 contexts × 5 concurrency levels
- Peak decode
- 626.7 tok/s
- Peak prefill
- 3,858 tok/s
- Peak decode
- 396.1 tok/s
- Peak prefill
- 3,709 tok/s
- Peak decode
- 760.9 tok/s
- Peak prefill
- 6,326 tok/s
- Qwen3.8-Flash-Next-W4A8-Paiton-RDNA4Radeon AI Pro R9700 32GB ×2 · vllm 0.29.0 · W4A81 context × 2 concurrency levels
- Peak decode
- 678.2 tok/s
- Peak prefill
- 7,121 tok/s
- Qwen3.8-Flash-Next-W3A8-Paiton-RDNA4Radeon AI Pro R9700 32GB ×2 · vllm 0.29.0 · W3A81 context × 2 concurrency levels
- Peak decode
- 745.2 tok/s
- Peak prefill
- 8,292 tok/s
- Qwen3.8-Flash-Next-W4A16Intel Arc Pro B65 32GB ×4 · vllm 0.30.0 · INT4_W4A166 contexts × 3 concurrency levels
- Peak decode
- 243.2 tok/s
- Peak prefill
- 3,229 tok/s
- Qwen3.8-Flash-Next-W4A16Intel Arc Pro B65 32GB ×4 · vllm 0.30.0 · INT4_W4A166 contexts × 3 concurrency levels
- Peak decode
- 159.2 tok/s
- Peak prefill
- 3,513 tok/s