moonshotai/Kimi-K2.6 model card

Fetched 2026-10-18. Weights repo created 2026-04-14 (HF API createdAt 2026-04-14T04:23:36Z, commit 7eb5002f6aad). Card: Kimi K2.6, open-source native multimodal agentic model from Kimi K2.5, 1T total and 32B activated, 256K context, Moonshot License. The benchmark table is a vendor claim, thinking mode on, temperature 1.0, top_p 1.0, context 262144, coding scores averaged over 10 runs on an in-house SWE-agent fork: SWE-bench Verified 80.2, SWE-bench Multilingual 76.7, SWE-bench Pro 58.6, Terminal-Bench 2.0 (Terminus-2) 66.7, LiveCodeBench v6 89.6, GPQA-Diamond 90.5, HLE-Full with tools 54.0, HLE-Full without tools 34.7 (text-only subset 36.4 without tools, 55.5 with tools), AIME 2026 96.4, HMMT 2026 Feb 92.7. The card publishes no MMLU-Pro number. The card K2.5 column is SWE-bench Verified 76.8, LiveCodeBench v6 85.0, GPQA-Diamond 87.6 and AIME 2026 95.8. The deployment section names vLLM, SGLang and KTransformers and says the K2.5 deployment method is reused. Weights are native INT4 (compressed-tensors, group size 32, pack-quantized), with attention, the shared expert, the dense MLP, the lm_head and the vision tower left out of the quant. HF safetensors metadata reports total 1,026,879,376,368 parameters, of which 1,014,686,023,680 are stored as I32 (INT4 packed into int32) and 12,193,329,648 as BF16.