Moonshot AI has published the full weights of Kimi K3.1 Pro, its most capable open-weight model, on Hugging Face. The repository, moonshotai/Kimi-K3.1-Pro, is now public and downloadable, with 96 weight files, configuration files and a licence, after independent checks on the morning of November 4 still found it unavailable. The release delivers on the promise Moonshot made at the model’s October 23 launch: that the full weights would follow by November 4.
What is in the release
The model card describes a mixture-of-experts system with 2.8 trillion total parameters and 104 billion activated per token, across 93 layers and 896 experts, with 16 experts selected for each token. It carries a one-million-token context window and accepts text and images, using a 401-million-parameter vision encoder. Weights ship in MXFP4 precision.
Moonshot’s own benchmark table puts Kimi K3.1 Pro at 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2%. It also reports 70.0% on HLE-Full with tools and 92.1% on BrowseComp, results behind its fourth-place Artificial Analysis ranking. All are company-reported scores; independent reproductions will be the real test.
For those who would rather rent than run it, the hosted route is unchanged. The model remains available through the Kimi API as kimi-k3.1-pro, at $2 per million input tokens and $10 per million output tokens, alongside a faster Highspeed variant. Moonshot recommends vLLM, SGLang and TokenSpeed for self-hosting, expecting serious operators to pick the weights up first.
The licence and the catch
The repository ships under a bespoke Kimi K3.1-Pro Licence, copyright 2026 Moonshot AI, which grants use, copying, modification and distribution subject to conditions. Its terms include provisions for model-as-a-service offerings, so anyone planning commercial hosting should read the licence itself rather than assume it matches earlier Kimi releases.
Size is the other catch. Independent estimates put the native checkpoint at roughly 1.5 terabytes, far beyond any single workstation or laptop. Running it properly means a cluster with high-memory accelerators and fast interconnects, plus the engineering to serve 104 billion active parameters at reasonable speed. Downloading the model is free; running it well is a serious infrastructure project.
Why It Matters
This is one of the largest open-weight releases ever, and it puts a near-frontier coding and agentic model into the hands of any lab, startup or researcher willing to host it. Open weights let others inspect, fine-tune and build on the model without depending on Moonshot’s API, and let researchers verify vendor claims first-hand.
The release also carries a safety trade-off Moonshot disclosed. The company removed cyber-related data from mid-training and reinforcement learning for the open-weight version, so the public model would not extend the frontier of offensive cyber capability. Its own test shows the cost: a checkpoint without the removal scored 90.5% on a hard cyber benchmark, against 81.0% for the released version.
Our read
Our read is that deadline-day delivery matters as much as the model itself. Moonshot promised the weights by November 4, and opened the repository on schedule after morning checks still found it closed. The open-weight race has a clear shape: Chinese labs, including Moonshot and Tencent with Hy4, publish near-frontier weights while leading American labs keep their best models closed.
What to watch is what happens next, not the launch table. Community quantisations will decide whether smaller labs can run the model at all, fine-tunes will show whether it holds up outside Moonshot’s harness, and independent benchmarks will confirm or dent the company’s scores.
Background
Kimi K3.1 Pro launched on October 23 as a post-trained version of Kimi K3, with larger post-training compute, and the launch strained Moonshot’s systems, forcing a quota reset. Moonshot described it as its most capable 3T-class open-weight model to date, and the weights were the second half of the announcement. That half has now arrived, as the launch post promised.
Source: Hugging Face model card and Moonshot launch post.
0 Comments