Kimi K3

What it is

Kimi K3 is the headline model of mid-July 2026 and, on its maker’s account, the largest open-weight model yet built. Moonshot AI announced kimi k3 on 16 July and released it on 17 July, days before the World AI Conference in Shanghai, timing that put a Chinese open-weight model at the front of the week’s news. Moonshot puts the model at 2.8 trillion parameters, though some reports place the figure nearer 2.7 trillion, and has scheduled the full weights for release on 27 July.

Features and architecture

Beyond raw size, kimi k3 carries the features expected of a current frontier model: a one-million-token context window, native vision so it can read images as well as text, and an always-on “thinking mode” that has the model reason before it answers. A context window that large lets it hold very long documents or codebases in a single session, while native vision means images are handled directly rather than described to it. The architecture leans on two techniques Moonshot had previously published as open research: Kimi Delta Attention, a hybrid linear-attention scheme, and Attention Residuals. Building the flagship on openly documented methods fits the open-weight positioning, since the company is competing in the open rather than behind a wall of secret tricks.

How it compares

On performance, the natural question is kimi k3 vs gpt-5.6, and here Moonshot is candid. Its own charts place kimi k3 behind only Claude Fable 5 and GPT-5.6 Sol overall, while beating Claude Opus 4.8 and GPT-5.5 on some coding and agentic benchmarks. Moonshot stops short of claiming the top spot and instead claims a place just below the two leading US flagships, which for an open-weight model is a strong showing. These are the company’s figures, so independent testing will be the real test.

Pricing and adoption

Pricing sits in an interesting middle. At $3 per million input tokens and $15 per million output tokens, kimi k3 is expensive by Chinese standards yet cheap next to the US flagships it trails. The model is compatible with the OpenAI SDK, which lowers the switching cost for developers already building against that interface and makes trying kimi k3 a small change rather than a rewrite. That combination of near-frontier quality, open weights and familiar tooling is what makes the release awkward for the incumbents.

Moonshot AI comes to this from a position of strength. The company raised $2bn in May 2026 at a valuation above $20bn and is backed by Alibaba. Its models are already in commercial use: Cursor drew on Kimi to help build its Composer 2 coding tool, and DoorDash delegates lower-level work to the earlier K2.6. That real-world adoption gives kimi k3 a credibility that benchmark charts alone cannot, and it explains why a Chinese open-weight release now commands attention across the field. The timing was no accident either: releasing days before the World AI Conference in Shanghai guaranteed kimi k3 a prominent hearing among the researchers, officials and investors gathered there, and for Moonshot a well-received flagship at that moment is worth as much as any leaderboard placement.

What to watch

Where kimi k3 sits in the wider field is at the sharp end of the open-weight challenge to the closed US labs. By landing a model just behind the leaders, releasing the weights and pricing below the flagships, Moonshot is testing whether openness plus value can erode the premium that GPT-5.6 and Claude Fable 5 command. For a Chinese lab to benchmark and price this close to the American leaders, and to give the weights away, is the sort of move that shifts expectations for the rest of the year. The full weights on 27 July, and the independent benchmarks that will follow, are the next things to watch. For more, see our large language models hub and wider AI coverage.