Kimi K2.7-Code: Moonshot open-weights a coding monster
Moonshot just dropped Kimi K2.7-Code, and the headline isn't that it's another trillion-parameter model. It's that they open-weighted it under a Modified MIT license and pointed the whole thing at one job: writing code with fewer tokens. Released June 12, weights live on Hugging Face, API at 95 cents in and 4 dollars out per million tokens, plus a 19-dollar-a-month CLI plan. You can also just host the thing yourself.
The numbers Moonshot reports over K2.6: +21.8% on its own Kimi Code Bench v2, +11% on Program Bench, +31.5% on MLS Bench Lite, and roughly 30% less reasoning-token burn for the same work. That last one matters more than the benchmark jumps. The whole game in coding agents right now is doing more inside a tighter token budget, because every wasted thinking token is money and latency. Same 1T architecture as K2.5 and K2.6 underneath, 32B active per token across 384 experts, 61 layers, MLA attention, 256K context, plus a MoonViT encoder so it can read screenshots and video.
Here's the bigger picture. This is the third Chinese-lab coding play in a week. Kimi Work shipped a local desktop agent, Xiaomi open-sourced MiMo Code, now Moonshot puts a frontier coding model on Hugging Face for anyone to run. The pattern is hard to miss: while the US labs guard weights and meter access, the Chinese labs are using open weights as distribution. Give away the model, win the developers, figure out monetization later.
Whether K2.7-Code actually beats Opus on real work is the open question, because vendor benchmarks always flatter the vendor, and we've watched independent evals like Endor on Fable 5 deflate exactly these claims. But open weights means you don't have to trust the chart. Download it, point it at your repo, see for yourself. Details at kimi.com, weights on Hugging Face.
← Back to all articles
The numbers Moonshot reports over K2.6: +21.8% on its own Kimi Code Bench v2, +11% on Program Bench, +31.5% on MLS Bench Lite, and roughly 30% less reasoning-token burn for the same work. That last one matters more than the benchmark jumps. The whole game in coding agents right now is doing more inside a tighter token budget, because every wasted thinking token is money and latency. Same 1T architecture as K2.5 and K2.6 underneath, 32B active per token across 384 experts, 61 layers, MLA attention, 256K context, plus a MoonViT encoder so it can read screenshots and video.
Here's the bigger picture. This is the third Chinese-lab coding play in a week. Kimi Work shipped a local desktop agent, Xiaomi open-sourced MiMo Code, now Moonshot puts a frontier coding model on Hugging Face for anyone to run. The pattern is hard to miss: while the US labs guard weights and meter access, the Chinese labs are using open weights as distribution. Give away the model, win the developers, figure out monetization later.
Whether K2.7-Code actually beats Opus on real work is the open question, because vendor benchmarks always flatter the vendor, and we've watched independent evals like Endor on Fable 5 deflate exactly these claims. But open weights means you don't have to trust the chart. Download it, point it at your repo, see for yourself. Details at kimi.com, weights on Hugging Face.
Comments