Moonshot AI published the full weights of Kimi K3.1 Pro on Hugging Face on 4 November 2026, as moonshotai/Kimi-K3.1-Pro — 96 files totalling about 1.56 TB, with no access gate. It is a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters per token, a 1-million-token context window and native multimodal input, making it the largest open-weight model release on the Hub to date. The release arrived on the date Moonshot promised when it launched the model on 23 October.

The benchmark numbers Moonshot reports are striking. In the company's own tests, K3.1 Pro scores 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4% — in other words, an open-weight model now sits within a point of the best closed coding model on a respected agentic benchmark. That is the headline for developers: frontier-level coding and agentic ability that can be downloaded, inspected and run on your own infrastructure, rather than rented through an API.

Two caveats keep the story honest. First, the model ships under Moonshot's own custom Kimi K3.1 Pro Licence — MIT-style, but with revenue and user-count thresholds — not a standard open-source licence, so commercial users should read the terms before building on it. Second, "open weights" does not mean "easy to run": at this size the model realistically needs around eight data-centre GPUs to self-host, so most people will still use it through a hosted provider rather than on their own hardware.

Even with those limits, this is one of the biggest open-model events of the year. A 1M-token context, native multimodality and near-frontier coding scores, freely downloadable, will reshape what smaller labs and independent developers can build — and it puts real pressure on closed-model providers to justify their prices. Watch for the first independent benchmark runs in the coming days: vendor-reported scores are a starting point, and the community's own tests will tell us how close K3.1 Pro really is to the frontier.

Sources