Moonshot

This is the biggest open model release in some time, and we covered it in a separate post and a podcast episode. It was released under a noncommercial license, requiring inference and fine-tuning providers to enter into a commercial agreement. Kevin Xu and Graham Webster argue in a post that these licenses enable potential future government action against US entities doing business with Chinese AI companies: But if a US company needs a contract with Moonshot to provide the inference tokens that Kimi K3 generates, the picture looks different. Some of the policy tools US officials and others have debated as potential levers to restrict Chinese open model use would more clearly apply.

An update to Kimi focusing a lot on token efficiency.

An update to the Kimi series, delivering stronger performance across the board and making it one of the best open models out there yet again. They also focus on long-horizon performance, showing that open models are capable of running over hours to complete tasks or optimize performance. Given the focus of everyone to build autoresearch-like systems, seeing open models catch up is important.

A continual pre-train on 15T tokens. Furthermore, this model is also multimodal! People on Twitter have replaced Claude 4.5 Opus with K2.5 for tasks that need a less capable but cheaper model. However, the writing capabilities that K2 and its successor were known for have suffered in favor of coding and agentic abilities.

The best open model, competitive with some of the best closed models. However, independent evaluation is a problem as third-party API providers struggle to implement the model correctly, something which we have seen, for example, with the release of GPT-OSS. As an example, running the agentic "Vending-Bench" from Andon Labs with a third party provider as opposed to the official API makes a huge difference: This is a huge problem plaguing open models. Moonshot also documents the tool calling accuracy in a repo, where a lot of providers perform sub-par, including vLLM with a schema accuracy <90%.

Similar to other (Chinese) labs, Kimi also experiments with attention variants.

An update to Kimi K2 focusing on its agentic abilities as well as coding.

It is hard to understate the impact of the model, which we've covered already. The new update is that they released a technical report today, with a bunch of nice methods, but nothing incredibly surprising. K2 has proven itself to be a capable model on various, unusual benchmarks, matching Opus on LMArena while becoming one of the most used models on OpenRouter.

An update to Kimi's visual reasoning MoE model. This version uses fewer thinking tokens, while having a superior performance over its predecessor.

A fine-tune of Qwen2.5 for agentic software tasks. The success of tools like Cursor or Claude Code shows the need for such models. For an analysis of SWE-bench Verified, the benchmark most of these models are evaluated on, see my (Florian's) analysis.

A vision MoE from Moonshot AI / Kimi, released under the MIT license. The performance of the model looks really solid, being on par with Qwen2.5 VL-7B, while using half the amount of activated parameters (but twice the total params). They also release a thinking version, boosting the scores even further. In their technical report, they state that they will train bigger models on more data. Given that Moonshot AI is one of the superstars in the Chinese LLM space, this team and model series is one to keep an eye on. It also continues the trend we observed: The big Chinese labs and companies continue to release very capable models under permissive OSS licenses, usually MIT or Apache 2.0.

Moonshot, mainly known for the Kimi models, has released a MoE to showcase the superiority of the Muon optimizer compared to AdamW.