MoE premier2.8T total params, Stable LatentMoE (16 of 896 experts active per token) built on Kimi Delta Attention (hybrid linear attention) + Attention Residuals. Gated MLA, Quantile Balancing routing, Per-Head Muon optimizer, SiTU activation; MXFP4 weights / MXFP8 activations (QAT). Native 1M context, native vision (text + image + video in, text out), 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 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 (GPQA-Diamond 93.5, HLE 43.5, Terminal Bench 2.1 88.3, FrontierSWE 81.2, MMMU-Pro 81.6) are Moonshot self-reported. Open weights announced (world's first open 3T-class model) but not yet public on HuggingFace as of release day - full weights scheduled by 2026-07-27. Cloud API $3/1M in, $15/1M out (cache-hit input $0.30/1M); self-hosting needs a 64+ accelerator supernode, so cloud-only for nearly everyone despite open weights.
- 2800.0B
- 1000k
- other
- Jul 2026
Scores #
Or run it in the cloud #
No per-token API provider pricing tracked for Kimi K3 yet.
For flagship list prices, see the
[token counter](/token-counter).
[Detailed API pricing page + JSON endpoint →](/models/kimi-k3/pricing)
Inference cost over time #
Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.