Moonshot AI has released Kimi K3, a 2.8-trillion-parameter multimodal reasoning model that approached or beat Claude Fable 5 and GPT-5.6 Sol on several coding and agentic evaluations published by the company. Moonshot described K3 as its most capable system and said its overall performance still trails those two proprietary models, making the launch a near-frontier result rather than a new leader across every category.
At maximum reasoning effort, K3 scored 91.2 on BrowseComp, compared with 90.4 for GPT-5.6 Sol and 88.0 for Claude Fable 5. It scored 42.0 on SWE Marathon, ahead of 39.0 for GPT-5.6 Sol and 35.0 for Fable, and led the same comparison group on Automation Bench. K3 also edged GPT-5.6 Sol on Program Bench and finished within half a point of it on Terminal Bench 2.1.
The results were less uniform on knowledge tests. K3 scored 93.5 on GPQA-Diamond, above Fable’s 92.6 but below GPT-5.6 Sol’s 94.1. On Humanity’s Last Exam without tools, its 43.5 trailed GPT-5.6 Sol at 44.5 and Fable at 53.3. That spread supports Moonshot’s claim of frontier-level performance without establishing K3 as the strongest general model.
The table is vendor-published and the comparisons are not fully standardized. Depending on the benchmark, models ran through KimiCode, Claude Code, or Codex, and some Fable results may include fallback to Claude Opus 4.8. Several Moonshot evaluations are internal. The full weights and technical report are also still pending, so independent evaluators cannot yet reproduce the complete performance picture.
K3 is a mixture-of-experts model that activates 16 of 896 experts. It combines Kimi Delta Attention with Attention Residuals, includes native vision, and supports up to 1 million tokens of context. Moonshot says those changes produced roughly 2.5 times better scaling efficiency than Kimi K2. The company recommends deployments with at least 64 accelerators, underscoring the infrastructure required to serve a model of this size even with sparse activation and low-precision weights.
The release therefore separates weight availability from practical deployability. Developers may soon be able to download the model, but running the full checkpoint will still demand infrastructure far beyond a typical enterprise server.
The model is available through Kimi’s web and mobile products, Kimi Code, and its API. Moonshot says it will release the full weights by July 27. That commitment lands as Beijing considers possible restrictions on overseas access to advanced Chinese models and as Chinese technology groups pursue new ways to narrow the AI hardware gap created by US export controls. If the release proceeds, K3 would pair near-frontier capability with downloadable weights at an unusually large scale.
–
By the Control Plane Editorial Team