Kimi-K3 model card (Section 3: Evaluation Results)

Read https://huggingface.co/moonshotai/Kimi-K3 for the benchmark figures on models/kimi-k3. As recorded by the reader: Read directly from the 'Evaluation Results' HTML table on the official HF card (column 'Kimi K3 (max)'). All results at reasoning effort 'max', temperature=1.0, top-p=0.95 for single-step tasks / 1.0 for agentic. Terminal-Bench 2.1 row evaluated with the Kimi Code harness (footnote 2). The card's other rows (CritPt, AA-LCR, HLE-Full, DeepSWE, ProgramBench, FrontierSWE, SWE-Marathon, PostTrainBench, MLS-Bench-Lite, SciCode, Kimi Code Bench 2.0, BrowseComp, DeepSearchQA, ResearchRubrics, GDPval-AA v2, Toolathlon-Verified, MCPMark-Verified, multimodal rows) are not in the allowed benchmark list and were excluded. The card publishes no AIME, MMLU-Pro, SWE-bench Verified/Pro, LiveCodeBench, IFEval, MATH-500, HumanEval or Arena-Hard scores.