Moonshot AI's 2.8-trillion-parameter flagship, ranked #1 on Arena's Frontend Code leaderboard. 1M context, native vision, open weights. Try it here, then open the full studio.
Moonshot shipped K3 on July 16, 2026. It's a mixture-of-experts design that routes each token through just 16 of 896 experts — roughly 1.8% of the pool.
Independent and self-reported figures side by side, including the places K3 comes second.
| Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 |
|---|---|---|---|---|
| Frontend Code Arena· blind developer votes | 1,679 | 2nd | — | — |
| GDPval-AA v2· 44 occupations, real tasks | 1,687 | 1,815 | 1,747.8 | — |
| Artificial Analysis Elo· long-horizon knowledge work | 1,547 | 1st | — | — |
| SWE Marathon· long-running engineering | 1st | — | — | — |
| Program Bench· program synthesis | 1st | — | — | — |
| Terminal-Bench 2.1· command-line agents | 2nd (−0.5) | — | 1st | — |
| Relative cost / task· Artificial Analysis est. | 1.0× | — | 1.1× | 1.9× |
| Context window· input tokens | 1M | — | — | — |
Figures from Moonshot's K3 announcement, Artificial Analysis, Arena.ai and Simon Willison's write-up. The honest picture: Moonshot's own benchmarks show K3 mostly losing to Claude Fable 5 and GPT-5.6 Sol overall. The cost row is a relative estimate normalised to K3.
896 expert subnetworks with only 16 active per token. Enormous capacity without paying full-dense compute on every request — the design that makes a model this size practical to serve.
In blind developer voting it ranked first at 1,679 points, ahead of Claude Fable 5. It also led SWE Marathon and Program Bench across Moonshot's coding suite.
Built for long-horizon coding and agent workloads. On BrowseComp, Moonshot notes that using the full 1M window rather than compaction at 300K lifts the score into the 90–91% range.
It reads images directly, and independent testing found the image understanding genuinely good — accurate, detailed descriptions from a rendered SVG, not vague summaries.
Two architectural additions — Kimi Delta Attention and Attention Residuals — which Moonshot credits for much faster long-context handling and better scaling efficiency than K2.
Moonshot committed to releasing full weights for research and commercial use, making K3 the largest open-weight model to date, taking that crown from DeepSeek's 1.6T v4 Pro.
The model is identical. What differs is everything around it.
K3 matters for a specific reason: it's the first Chinese open-weight release that competes with the top American systems on capability rather than on price. That reframes what an open model can be used for.
This is its clearest win. K3 ranks first on Arena.ai's Frontend Code leaderboard at 1,679 points, ahead of Claude Fable 5, in blind developer voting — a 17-place jump from K2.6 with a 76% pairwise win rate.
It led SWE Marathon and Program Bench, and trailed GPT-5.6 Sol on Terminal-Bench 2.1 by only half a point. Combined with the 1M-token window, that suits tasks which run for a long time and accumulate context.
Artificial Analysis put K3's overall Elo at 1,547 — a 732-point jump from K2.6, behind only Claude Fable 5. Estimated cost per task came in at roughly half that of Claude Opus 4.8, and Moonshot reports it uses 21% fewer output tokens than K2.6 on equivalent tasks.
Moonshot's own benchmarks show K3 losing to Claude Fable 5 and GPT-5.6 Sol overall. And it currently ships with only one reasoning effort level, “max” — it reasons heavily on everything. For quick, cheap questions that's the wrong tool.