Kimi K3

2.8T total params, Stable LatentMoE - 16 of 896 experts active per token. Built on Kimi Delta Attention (hybrid linear attention) plus Attention Residuals, with Gated MLA, Quantile Balancing routing, a Per-Head Muon optimizer, and SiTU activation. Weights are MXFP4 with MXFP8 activations (QAT).

• Context and I/O: native 1M context; native vision (text + image + video in, text out).
• Reasoning and tools: always-on thinking mode (reasoning_effort currently “max” only); full function/tool calling with MCP; automatic prompt caching (>90% cache-hit on coding).

Independent eval. #1 on the Artificial Analysis webdev arena (1679, ahead of Claude Fable 5 at 1631) and #4 of 189 on the AA Intelligence Index (57) - comparable to Opus 4.8 / GPT-5.5, behind Fable 5 / GPT-5.6 Sol.

Vendor benchmarks (Moonshot self-reported): GPQA-Diamond 93.5, HLE 43.5, Terminal Bench 2.1 88.3, FrontierSWE 81.2, MMMU-Pro 81.6.

Open weights shipped 2026-07-27 - the world’s first open 3T-class model, on HuggingFace at moonshotai/Kimi-K3. Full-precision self-hosting still needs a 64+ accelerator supernode. Unsloth’s 1-bit dynamic GGUF (UD-IQ1_S, unsloth/Kimi-K3-GGUF) shrinks the weights 62% to 594GB (~610GB to run in RAM) at ~78.9% top-1 agreement vs fp16 - it fits in RAM on a 748GB DGX Station, and on a Mac it loads via SSD memory-mapping with a ~128GB working set (slower, but it runs).

Cloud API: $3/1M in, $15/1M out (cache-hit input $0.30/1M). On Ollama Cloud, run ollama run kimi-k3:cloud - but K3 is not part of the flat subscription: it bills per-token from a separate metered “extra usage” balance, and a 402 “extra usage balance is empty” response means you must add funds or enable auto-reload at ollama.com/settings.

Agents on Rails benchmark (Aug 2026, Le Mans round). 90.5% accuracy on 63 runs - tied with Claude Fable 5 and 2 points behind Claude Opus 5 across the 16-model field. Cost was roughly half of Fable’s bill, making it the strongest accuracy-per-dollar among the top cluster.