Kimi K3.1 Pro launches at No. 4 on a third-party index; if its weights ship, the open-closed gap narrows to 3 points

Moonshot AI has released Kimi K3.1 Pro, the first model in a new series. The release confirms that "Kibitz," a model that appeared anonymously on Arena on Oct. 22, was Kimi K3.1. The model placed fourth on the Intelligence Index from third-party evaluator Artificial Analysis. Moonshot says it will release the weights by Nov. 4. If it keeps that promise, the gap between the top open-weight model and the top closed model on the third-party index would narrow from 12 points to 3.

The story is less about the performance numbers themselves than about whether those numbers end up on the open side. A single promised date could significantly reshape the industry's landscape.

Unmasking Kibitz: from leaks to official announcement

Several clues about Kimi K3.1 surfaced ahead of the official announcement. On Sept. 29, a "Kimi K3.1" identifier was spotted in Moonshot's API listings. On Oct. 21, screenshots circulated that purported to show the name in Kimi Code's model-switching menu. Then on Oct. 22, an anonymous model called "Kibitz" appeared on Code Arena. It identified itself as Kimi and, according to testers, used the same tokenizer as K3. It also matched the naming conventions Moonshot has used for its test models in the past.

At that point, Moonshot had confirmed nothing, and nothing had appeared on Hugging Face. The official blog post has now turned that speculation into fact. Unite.ai describes Pro as the first model in the K3.1 series, with the rest of the series still to come.

"Kalmia," which appeared on Arena the same day, is only believed by the community to be Qwen4; Alibaba has not confirmed it. Of the two mystery models, only Kibitz has been identified.

Fourth place, and the math from 12 points to 3

Kimi K3.1 Pro placed fourth on the Artificial Analysis Intelligence Index. Set against the existing rankings, that position stands out.

Its predecessor, Kimi K3, scored 44. As of Oct. 19, the top open-weight model was Xiaomi's MiMo-V2.6-Pro at 46, followed by DeepSeek V4.1-Flash-1017 at 45. The top closed model was Claude Opus 5.5 at 58, with Muse Wonder Full and Claude Sonnet 5.5 tied at 56. The difference between 46 and 58 is the 12 points in the headline.

If the K3.1 Pro weights are released, the open-weight lead would change hands, and the gap to Opus 5.5 would be 3 points. Working backward from those figures puts K3.1 Pro's score at roughly 55, which fits its fourth-place ranking behind Opus 5.5, Muse Wonder Full and Sonnet 5.5. That would be a gain of more than 10 points over its predecessor, K3.

A 3-point gap is also small compared with the gaps the evaluator has recorded before. NVIDIA's Nemotron 3.5 Super VL, whose weights were released on Oct. 21, scored 26 and was the top U.S. open-weight model. The top Chinese open-weight model would then sit roughly 30 points above it.

The Nov. 4 condition: promises kept, and promises pending

For now, however, K3.1 Pro is not an open-weight model. The 3-point gap only applies if the weights are released by Nov. 4. Until then, the top of the index is held by four closed models, one of which now happens to be Chinese.

Labs have mixed records on "weights coming later" promises. StepFun released the Step 5 Preview weights on Oct. 15, the date it had announced. Reflection AI released the Beam weights on Oct. 16, having promised them "this month." By contrast, the Mistral Large 4 weights have yet to appear on Hugging Face, after the company said on Oct. 6 they would come "by the end of the month." MiniMax, too, had said only that the M3.1 Flash weights would come "later," without giving a date, as of Oct. 19.

Moonshot has reason to keep its date. As Chinese outlets including Wallstreetcn have reported, Moonshot reportedly plans to list in Hong Kong in the January–March quarter of 2027. Its valuation in its last private round before the listing was reportedly about $50 billion (secondhand reporting citing Bloomberg; the company has not confirmed it). Open-weight releases are what built the Kimi name, so whether Moonshot meets its promised date is also a question of credibility with investors.

What the index doesn't measure, even as the gap narrows

It is too early to read a 3-point gap as "the gap between open and closed has disappeared." The index averages many benchmarks, and some differences don't show up in it.

In the third round of First Proof, whose results came out on Oct. 14, a system running on the predecessor Kimi K3 failed to solve any of 10 unpublished research-level math problems. GPT-6 Astra solved seven on its own. Mathematician Abouzaid said the gap between open and frontier models was "much larger than in other fields." Terence Tao has also written that an open-model platform offered to academics is about a year behind the frontier. How much of this gap K3.1 Pro has closed has not yet been measured.

A narrowing gap carries weight in another direction, too. Epoch AI estimated on Oct. 16 that open-weight models lag closed models by six to seven months in cyber capabilities. Anthropic reported in September that by modifying the weights of Z.ai's GLM-5.3, its safeguards could be bypassed 100% of the time. Once weights are released, the distributor cannot enforce a model's safeguards. If a model within 3 points on the index circulates as open weights, those lag estimates will likely shrink as well. How much Moonshot discloses about safety will need to be checked alongside the weight release.

What to watch next

First, whether the weights appear on Hugging Face by Nov. 4. When they do, it will be worth checking the license and whether they are the weights for Pro itself or for another model in the series. Second, how K3.1 Pro performs on third-party evaluations separate from the index, such as First Proof. Finally, whether Mistral Large 4 and MiniMax M3.1 Flash, whose weights have also been promised for around the same time, arrive, along with Arena's Kalmia (believed to be Qwen4). If Chinese labs keep releasing weights, the race for the open-weight lead could reach just below the closed models on the index by early November.