Moonshot AI Schedules Weight Release Without Repository Documentation

Moonshot AI has set a target date of Nov. 4, 2026, to distribute the full model weights for its latest release, though crucial governance terms remain unconfirmed. Enterprise deployers looking to integrate the Kimi K3.1 Pro model are being advised to closely monitor licensing documentation rather than relying strictly on the developer calendar.

A direct audit of Moonshot’s official Hugging Face organization reveals 19 existing repositories, with the most recent entry being Kimi-K3 from Sept. 2. As of now, no repository, model card, or licensing file has been created for the upcoming release. Without a published repository, the precise operational terms for commercial usage remain unconfirmed.

Commercial Thresholds and the Club of Four Standard

In the absence of a distinct license for the new release, the existing Kimi K3 License serves as the sole precedent for enterprise users. That agreement mandates that any business operating a Model-as-a-Service model with aggregate revenue exceeding $20 million over a 12-month period must secure a separate contract with Moonshot AI.

This $20 million revenue ceiling forms part of the broader Club of Four convention, an informal alignment among AI developers including StepFun, Meta, Moonshot AI, and Mistral. Moonshot pioneered this structure, calculating the gate using a rolling 12-month aggregate of group revenue across all operational sources.

The approach stands in contrast to recent industry missteps, such as StepFun failing to publish its repository and license past an Oct. 15 target. Meanwhile, Mistral's ongoing countdown for Mistral Large 4 continues to test whether the $20 million revenue gate will harden into a firm sector standard.

Technical Capabilities and Benchmark Performance of the Kimi K3.1 Pro Model

Described as a post-trained iteration of Kimi K3, the new release is classified as a 3T-class open-weight model featuring a 1-million-token context window. The system is already active across Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, alongside a Highspeed variant designed to operate two to three times faster.

To mitigate security risks, Moonshot reported removing cyber-related data from both training phases prior to releasing the open weights. On the internal xdev-bench-hard benchmark, an unstripped checkpoint scored 90.5%, whereas the final release candidate achieved 81.0%.

Vendor-reported benchmarks place the model at 65.7% on Terminal-Bench 4.0, trailing slightly behind Anthropic’s Claude Opus 5.5 score of 66.4%. Moonshot emphasized that the model's cyber capabilities intentionally remain behind proprietary frontier models that have been available for at least three months, sticking to a commitment not to push the frontier of cyber capability in open releases.