Kimi K3.1 Pro Open Weights Land: 2.8T Params, 1.56 TB Download
3 min readMoonshot AI has followed through on its promise. The Kimi K3.1 Pro open weights are now listed on Hugging Face under the repository moonshotai/Kimi-K3.1-Pro, according to a model-card breakdown published by Shattered on November 5 that cites TechNode’s reporting. The listing confirms a 2.8 trillion parameter mixture-of-experts model with a one million token context window, released as a 1.56 TB download.
Background: the November 4 promise
When Moonshot launched Kimi K3.1 Pro on October 23, it published self-reported benchmarks that put the model within a point of Claude Opus 5.5 on Terminal-Bench 4.0 and ahead of GPT-6 Astra on SWE-Marathon, cut API prices to $2 and $10 per million tokens, and said the weights would be released by November 4. It also said it had stripped cyber data from mid-training and reinforcement learning before the open release, accepting a drop from 90.5 to 81.0 percent on an internal exploit benchmark in exchange for a safer public model. AI Incider covered that launch at the time.
What the Kimi K3.1 Pro open weights contain
The model card lists 2.8 trillion total parameters with 104 billion active per token, routed across 896 experts of which 16 fire on each token. The vocabulary holds 160,000 entries, and a 401 million parameter MoonViT-V2 encoder handles image input alongside text. Weights ship in MXFP4 with MXFP8 activations, which is why a 2.8 trillion parameter model fits in 1.56 TB; a 16-bit copy would run to roughly 5.6 TB. The model is post-trained from Kimi K3, and the context length is 1,048,576 tokens.
The release carries a custom Kimi K3.1 Pro License rather than Apache 2.0 or MIT. The terms had not been analyzed in detail at the time of writing, so questions about commercial use, redistribution of fine-tunes and training on outputs remain open. Moonshot has not issued an on-record statement about the release, and no independent benchmarks of the open checkpoint have appeared yet.
Why it matters
At 2.8 trillion parameters, this is by far the largest open-weight model available, roughly 3.6 times the size of Tencent’s 770 billion parameter Hunyuan Hy4 released this week. Few organizations can host it: even at 4-bit precision the weights need well over a terabyte of accelerator memory spread across a multi-node cluster. Expect hosted endpoints from cloud providers to arrive before most teams run it themselves, followed by distilled and pruned derivatives. The bigger signal is strategic. A Chinese lab has put frontier-class weights, trimmed of their sharpest cyber skills, in public hands within two weeks of launch, which keeps the pressure on US labs moving the other way with gated releases such as this week’s GPT-6 Cyber. Watch for the license terms, the first third-party evaluations and the first serving providers.