Moonshot's Kimi K2.7 Code Sharpens Its Open-Source Edge

Moonshot AI has put out another coding model, and this one is free to download. On 12 June the Beijing lab released Kimi K2.7 Code, the latest in a series that has shipped at a remarkable pace, with weights published on Hugging Face under a modified MIT licence. For developers who would rather run their own model than rent one, that licence is the headline: the system can be downloaded, inspected, and self-hosted rather than reached only through an API.

Under the hood it keeps the same broad shape as recent Kimi releases. It is a mixture-of-experts model with roughly a trillion parameters in total, but only about 32 billion of them fire for any given token, spread across 384 expert networks, with a context window of 256,000 tokens. That design is what lets a very large model run at a more manageable cost, since most of the network stays idle on any single request.

Faster thinking, not just bigger

The interesting claim this time is about efficiency rather than raw size. Moonshot reports that K2.7 Code uses around 30 percent fewer reasoning tokens than its predecessor while scoring higher on coding tests, with gains it puts at 21.8 percent on its own Kimi Code Bench v2 and 31.5 percent on a separate agentic benchmark. Fewer reasoning tokens matters because that hidden "thinking" is something you pay for, in money on an API and in time when you are waiting for an answer. Doing more with less of it is a real practical win, if the numbers hold up outside the lab's own testing.

As always with self-reported benchmarks, the sensible move is to treat them as a starting point rather than a verdict. According to Crypto Briefing and Digital Applied, the model is pitched squarely at coding agents, the kind of tooling that plans and edits across a whole project rather than answering one question at a time. Independent testing over the coming weeks will settle how it stacks up against the proprietary leaders on real work.

The open-weight pressure keeps building

For those who would rather not host it themselves, Moonshot offers the model through its own API at $0.95 per million input tokens and $4.00 per million output tokens, along with a flat monthly plan for command-line use. Either way, the release adds to a steady stream of capable open-weight models coming out of Chinese labs, the same wave that has carried MiniMax and others toward the frontier.

Moonshot, which has been reported to be chasing a $30 billion valuation, is clearly betting that giving the weights away builds the kind of developer goodwill that money cannot. Every strong open release also chips away at the pricing power of the closed labs, and that may be the most consequential part of the story. The competition is no longer only about who has the smartest model. It is about how cheaply you can get a good one.