Kimi K3.1 Pro hits 61.2% on long-horizon code bench, up from 55.95%
vllm_project · x · 2026-11-05
K3.1 Pro scores 61.2% on the internal Kimi Code Bench 3.0 Long-horizon, up from K3's 55.95%. The benchmark covers multi-step execution, tool use and end-to-end engineering, with both results using the Claude Code harness at maximum reasoning effort.
Serving work around it includes token-sharded prefill, exact top-k merging for decode context parallelism, and a fused low-latency decode path, letting K3.1's DSA sparse attention plus recurrent Kimi Delta Attention keep MLA latents, indexer keys and KDA state in one packed cache.
Related event: vLLM Ships Day-0 Support for Kimi K3.1 Across Training and Serving(5 posts)→
More from Infra
- LFM3-Omni-2B ported to run fully offline on iPhone at ~80 tok/s via custom WebGPU kernels — victormustar · 2026-11-05
- Solo dev's open-source browser agent uses local models to watch your screen and trigger actions — Roy3838 · 2026-11-05
- LFM3-Omni-2B runs fully offline on iPhone at ~80 tok/s via custom WebGPU kernels — xenovatech · 2026-11-05
- Mistral Large 4 1T hits HF Inference Endpoints at $10/hr on 4x A100 — ngxson · 2026-11-05
- NVIDIA accelerates structural simulation in Abaqus with CUDA-X — NVIDIAAI · 2026-11-05
- SpecAgent: multi-session Claude Code pipeline with measured 3x context cost savings — AdOtherwise9458 · 2026-11-05