Weights are coming on 4 November. The governance choice landed with the launch blog: Moonshot will ship Kimi K3.1 Pro open, after scrubbing cyber-related training signal so the download sits below the cyber frontier its own pipeline could have produced.
That is the story worth filing under open-weight governance. Capability marketing (3T-class, million-token context, Highspeed at two to three times the base model's speed) matters for buyers. The cyber curb matters for everyone who will run the checkpoint without Moonshot in the loop.
K3.1 Pro launches with a measured cyber cut and a November weight date
Moonshot introduced Kimi K3.1 Pro as the first model in its K3.1 series and its most capable 3T-class open-weight model to date, with a 1-million-token context window. Kimi K3.1 Pro Highspeed is positioned at roughly 2-3× the speed of K3.1 Pro. Both are available on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API; full model weights are scheduled by 4 November 2026.
On cybersecurity, the blog says vulnerability-research performance rose with post-training scale because cyber content sat in the general mid-training corpus and cyber tasks sat in the RL mix. For the open-weight release, Moonshot identified and removed cyber-related data from both stages while keeping general-purpose coding tasks. On in-house xdev-bench-hard (expert-curated source-code vulnerability discovery and exploitation tasks with sandboxed arbitrary-code-execution checks), the checkpoint without those risk-reduction measures scores 90.5% pass@3; final K3.1 Pro scores 81.0%. Reuters reported the curb on 24 October 2026, noting that the comparison is company-reported and independently unverified.
Moonshot binds the weight drop; downstream hosts inherit the residual risk
Moonshot AI ships the product routes today and the weight files by the stated November date. Enterprise administrators can enable K3.1 Pro for organisations; access is off by default at launch. Anyone who downloads, fine-tunes, or hosts the open weights inherits residual dual-use capability that remains after scrubbing; Moonshot itself says stronger general coding still generalises into cyber tasks. Defensive security teams hoping for open auditors and attackers hoping for local tools read the same checkpoint.
A public scrub delta on the path to indiscriminate weights
Publishing an 81.0 versus 90.5 internal cyber-bench delta as part of an open-weight launch narrative is the governance move. Moonshot states two principles: avoid releasing open weights that extend the absolute cyber frontier, and progressively raise the cyber capability available in open weights so defensive users can audit code on infrastructure they control. Vendor coding claims include Terminal-Bench 4.0 at 65.7% with a Claude-Code harness, narrowly behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2% on the blog's cited comparison; treat those figures as vendor-reported pending harness-matched third-party runs.
Vendor benches, internal cyber eval, and post-download control limits
The product launched with the K3.1 Pro blog; Reuters covered the cyber curb on 24 October 2026; weights are promised by 4 November 2026. Cyber numbers are pass@3 on Moonshot's xdev-bench-hard, without hard timeouts or output token limits in the described protocol. Terminal-Bench and other leaderboard citations use stated harnesses and mixed sources; they are marketing evidence until reproduced. Once weights circulate, Moonshot cannot revoke copies the way a hosted API can revoke keys. China's draft export discussion of AI malware and vulnerability-discovery tools, noted in Reuters coverage, is policy context around the launch rather than a licence term on the checkpoint.
Inventory open-weight cyber posture before 4 November
Security and model-risk teams should record the 81 versus 90.5 scrub delta next to any plan to mirror K3.1 Pro weights, and decide whether local fine-tuning could undo the curb. Procurement comparing Terminal-Bench screenshots should demand harness-matched evals against Opus 5.5 and Astra on their own tasks. Open-source defenders who want the checkpoint for code audit can plan hardware for a 3T-class, 1M-context serve without waiting for API quotas, while accepting that the same files leave Moonshot's trust boundary.