Moonshot AI releases Kimi K3.1 Pro weights, a 2.8-trillion-parameter open-weight model
China's Moonshot AI has published the full weights of Kimi K3.1 Pro on Hugging Face, giving researchers and developers free access to a 2.8-trillion-parameter multimodal model with a one-million-token context window.
Moonshot AI has released the full open weights of Kimi K3.1 Pro, its most capable model to date, fulfilling a promise made at the model's launch to make the weights freely downloadable by 4 November 2026.
Kimi K3.1 Pro was launched on 23 October 2026 as the first model in Moonshot's K3.1 series, a post-trained successor to Kimi K3, and has been available since then through the Kimi app, Kimi Work, Kimi Code and the company's application programming interface. The weight release, published on Hugging Face as moonshotai/Kimi-K3.1-Pro, is what turns a hosted product into a model that independent teams can download, inspect, run and build on themselves.
The scale is formidable. The model card lists 2.8 trillion parameters in total, with 104 billion active for each token, routed through a mixture-of-experts design with 896 experts of which 16 are selected per token. It is natively multimodal, taking text and images, and keeps the one-million-token context window, 1,048,576 tokens, that has become Kimi's signature, allowing very long documents and codebases to be held in a single conversation.
Under the hood, Moonshot made a targeted architectural change: in 24 of the model's 93 layers, full gated multi-head latent attention is replaced by sparse gated DeepSeek Sparse Attention, with a top-2048 indexer, a change designed to speed up processing across those very long contexts. A 401-million-parameter MoonViT-V2 vision encoder handles the image side of the model.
The download itself underlines the gap between open weights and easy access. The weights occupy about 1.56 terabytes across 96 files, stored at 4-bit MXFP4 precision after quantisation-aware training, and Moonshot's own deployment recipes call for a data-centre node, such as two Nvidia GB300 trays, one B300 node or one AMD MI355X node, running eight or more high-end graphics processors. For most teams, the hosted API will remain the practical route.
That API prices Kimi K3.1 Pro at $2 per million input tokens and $10 per million output tokens, with cached input at $0.20, down from $3 and $15 for Kimi K3, alongside a faster Highspeed variant at higher rates. The licence follows the Kimi K3 pattern: free to use, change and sell, but a company running a model-as-a-service business with more than $20 million in revenue over 12 months must sign a separate agreement with Moonshot, and products above 100 million monthly users or $20 million in monthly revenue must display the model's name.
On performance, Moonshot's own figures put Kimi K3.1 Pro at 65.7 per cent on Terminal-Bench 4.0, a test of coding agents on real terminal tasks, against 66.4 per cent for Anthropic's Claude Opus 5.5 and 58.2 per cent for OpenAI's GPT-6 Astra, with further vendor-reported scores of 94.2 on GPQA-Diamond and 79.6 on OSWorld. Those numbers have not been independently reproduced, and benchmark presentation in this fast-moving field is routinely contested, so they should be read as the company's claims rather than settled fact.
Moonshot also says it took a deliberate safety step before release, tracing rising exploit skill in the model to cyber content in its general mid-training corpus and reinforcement-learning environments and removing both, a reminder that the same agentic coding ability that makes the model useful also raises dual-use questions that open-weight releases cannot fully control once the files are public.
Deployment recipes for vLLM and SGLang were published alongside the weights, and 18 providers already listed the model in catalogues by 4 November, suggesting that hosted access will spread quickly even for teams that never download the full 1.56 terabytes themselves.
What happens next will test whether open weights at this scale genuinely democratise frontier capability or simply move the bottleneck. Researchers can now probe, fine-tune and stress-test a 3-trillion-class model without Moonshot's permission, but only those with data-centre hardware, or the budget for hosted inference, can run it at full strength, and the race with other open-weight laboratories, and with closed rivals, will be judged on independent benchmarks and real-world reliability rather than launch-day claims.
More in Technology & AI
Photo: Programming code on a screen (file image)
Technology & AIOpenAI to acquire Pierre Computer and fold its Code Storage system into Codex
OpenAI has announced plans to acquire Pierre Computer Company, the developer-tools startup whose Git-based Code Storage service and open-source Trees and Diffs libraries already power parts of its Codex coding app.
Photo: Meta headquarters in Menlo Park, California (file image)
Technology & AIMeta acquires AI coding startup Conductor and winds down its cloud service
Meta has acquired Conductor, the startup whose app runs teams of AI coding agents in parallel, with the team joining Meta Superintelligence Labs and Conductor Cloud shutting down on 20 November.
Photo: A silicon wafer used in semiconductor chip manufacturing (file image)
Technology & AIEU to invest €525 million in chip startups in bid to close the gap with the US and China
The European Union will put €525 million into chip startups, including €440 million in grants and €85 million for a new European Chips Design Platform, as Brussels tries to strengthen a sector where its design firms hold less than 1 per cent of global revenue.