Moonshot publishes Kimi K3.1 Pro weights: 2.8 trillion parameters under a custom licence
The Chinese lab released full weights for a mixture-of-experts model with 104 billion active parameters, a 1M-token context and image input. It beats GPT-6 Astra on several coding benchmarks but trails Claude Opus 5.5 on the hardest tasks.
What is inside
Kimi K3.1 Pro is a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active per token, built on 93 layers. It offers a context window of 1,048,576 tokens, accepts text and images, and exposes three reasoning levels: low, high and max.
Moonshot announced the model on 23 October and published the full weights on 4 November under the Kimi K3.1 Pro License, a custom licence rather than a standard open-source one. Self-hosting is a serious infrastructure project: the weights are quantised to MXFP4 and the model needs a large GPU cluster. Moonshot recommends vLLM, SGLang or TokenSpeed.
Benchmarks: close to the top, not everywhere
On Terminal-Bench 4.0 it scores 65.7%, ahead of GPT-6 Astra (58.2%) and Claude Fable 5.1 (57.9%), just behind Claude Opus 5.5 (66.4%). It leads the comparison on SWE Marathon v1.1 with 51.7% and on HLE-Full with tools with 70.0%.
Agentic results are strong too: 92.1% on BrowseComp, 90.0% on MMMU-Pro with Python and 79.6% on OSWorld 2.1. The gaps appear on the hardest reasoning and programming tasks: 31.5% on Program Bench against 65.0% for Opus 5.5, and 39.4% on HLE-Diamond against 65.8% for Astra.
Moonshot also removed cyberattack-related training material before release. On an internal exploitation test the pruned model solved 81% of tasks, down from 90.5% for an unpruned version but still far above Kimi K3 at 33.3%. The model is available through Moonshot's apps and API, with a Highspeed variant that runs two to three times faster.
What open weights mean in practice
For most companies the licence matters less than hosting. A 2.8-trillion-parameter model is not something a normal team runs on-premises. The realistic options are Moonshot's API, a Western cloud that hosts it, or a smaller distilled model later.
Where the weights do matter is data sovereignty. Banks, hospitals and public-sector teams that cannot send data to a foreign API can, in principle, run the model in their own environment. That is a genuine alternative, but one that requires GPU capacity, MLOps skills and a legal review of the custom licence.
Sources
More news
DeepSeek open-sources an autonomous coding agent
The Chinese lab published a prototype autonomous programming agent on GitHub — an open answer to Anthropic's Claude Code.
Read articleTencent unveils Hy4: an open model trained only on defensive cybersecurity
The Chinese giant showed a model with vision, computer use and stronger agentic skills. It excluded exploitation material from training and promises weights in the coming days.
Read article