Morning Edition · Wednesday, October 28, 2026Published at 2:21 AM EDT · New York
Moonshot's own benchmark tables put the 2.8-trillion-parameter mixture-of-experts model ahead of Claude Opus 5.5 on Humanity's Last Exam, but independent scoring ranks it fourth.

Moonshot AI released Kimi K3.1 Pro on October 23 through its own services. It has promised the full weights by November 4 under a license close to MIT, with added conditions for very large companies, DataCamp reported. The model is a mixture-of-experts design with 2.8 trillion total and 104 billion active parameters, a 1-million-token context window and multimodal input. If the checkpoint ships in the same compressed numerical format as K3, called MXFP4, it will take up about 1.56 TB on disk, according to Yotta Labs.
Moonshot's published comparisons use Claude Opus 5.5 as the main baseline. With tools, Kimi K3.1 Pro scores 70% on Humanity's Last Exam, against 67.7% for Opus 5.5 and 57.2% for GPT-6 Astra. It scores 65.7 on Terminal-Bench 4.0, against 66.4 for Opus 5.5, per MindStudio's review. These are vendor-reported numbers. The independent Artificial Analysis Intelligence Index scores the model at 55, placing it fourth, behind Opus 5.5 at 58 and level with Claude Sonnet 5.5 and Meta's Muse Wonder Full at 56. The gap between the two sets of numbers reflects the normal difference between a lab's chosen benchmarks and a fixed third-party suite.
The more consequential detail is a safety change. Moonshot deliberately reduced the model's offensive-security skills before releasing the weights, Technology.org reported on October 26. A downloadable checkpoint cannot be recalled or monitored once it is out. Whatever capability ships in the weights becomes available to anyone who fine-tunes it, which is why removing that capability before release is the only control an open-weight lab retains. Whether the removal survives fine-tuning is an open question that no independent group has yet tested.
The release comes as China's hardware base shifts. Nvidia's share of China's data-center AI chip market fell to 34% in the first half of 2026. Nvidia shipped 678,000 accelerators, down 44% year over year, while Huawei shipped 572,000 for a 28% share, according to an IDC report covered by Yahoo Finance. Beijing is also drafting its own export-control list, adding about 40 items across AI, solar power and biomedicine.
Part of a tracked trend
Chinese Open-Weight Models Emerge as the Non-US AI Stack
As Washington restricts foreign access to US frontier models, governments and enterprises cut off from American AI increasingly standardize on downloadable Chinese open-weight models, splitting the world into competing AI supply blocs rather than a single frontier.
Start a discussion in Townsquare.
More from this edition
Moonshot, Huawei and inference hosts outside the US gain customers whom US model-access rules exclude, plus a safety defense against Western bans on hosting the weights. US closed-API vendors lose pricing power with price-sensitive buyers.
Moonshot did remove cyber training data, per Notebookcheck. Its only published comparison shows a modest effect: the filtered model still solved 81.0% of an unpublished in-house exploit test, against 90.5% for the unfiltered build, Implicator reported. Beijing's October 16 draft export catalogue would restrict the export of model weights above about 2 trillion parameters, so if it is adopted after comments close on November 15, the 2.8-trillion-parameter release could fall under it.
An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.
What this means
A near-frontier model with downloadable weights reaches every government and company that US model-access restrictions exclude, and it reaches them at the cost of hosting rather than at API margin. The exposed parties are closed API vendors that sell to price-sensitive and non-US buyers, and US policymakers who treat model access as a lever of control. The benchmark claim is already discounted, since the independent index ranks the model fourth. What is new is a Chinese lab removing cyber capability before release. That gives Moonshot a defense against the argument that Chinese open weights pose an uncontrolled security risk, an argument that could otherwise justify US or EU restrictions on hosting them. The precedent is DeepSeek-R1 in January 2025. Its open release led to rapid adoption by Western cloud hosts, followed by government bans on its consumer app rather than on the underlying weights. The likelier outcome is that major non-Chinese inference hosts list K3.1 Pro within weeks of the weights appearing, and that the US debate shifts toward restricting the hosting of Chinese weights inside the US. This depends on Moonshot shipping the weights on schedule. A delay past November 4, or an independent red-team test showing the cyber capability returns after light fine-tuning, would prove the read wrong.
What to watch
Observations to monitor, not financial advice.
Source: Anthropic News
Comments
0No comments yet.