The Beijing lab behind Kimi put out its largest open-weight model yet on October 23, a refresh of K3 called K3.1 Pro, plus a faster sibling. Moonshot AI says the pair carries a 1M-token context and 2.8T total parameters with 104B active at a time.
Both are post-trained versions of K3 rather than new pretrains. The company reports spending substantially more compute on post-training, and it optimized the stack to run joint reinforcement learning across hundreds of downstream domains at once.
The Highspeed variant is pitched at two to three times the throughput of the standard model with the same capabilities, a split that matters most for agent loops that call a model hundreds of times per task. Both ship through the Kimi website, the mobile apps, the Kimi Work desktop client, and the paid coding tiers.
On Terminal-Bench 4.0 the company reports 65.7%, above GPT-6 Astra at 58.2% and a shade under Claude Opus 5.5 at 66.4%. It also claims 64.4 on OfficeQA Pro v2, a document-heavy finance test, and gains of 7.2 points on financial modeling and 11.6 on financial research against K3.
Moonshot attached a cyber capability note to the release. It says a checkpoint trained without risk-reduction measures solved 90.5% of its internal vulnerability-discovery benchmark at pass@3, while the shipped model reaches 81.0%. The lab says it will not push the absolute frontier of offensive cyber capability into open weights, but intends to keep raising what it publishes. Full weights are due by November 4.
