Moonshot AI announces Kimi K3.1 Pro with 1M-token context; weights due Nov. 4

Moonshot AI has announced Kimi K3.1 Pro, a multimodal reasoning model with a 1-million-token context window, which the company calls its most capable 3T-class open-weight model to date. A faster companion, Kimi K3.1 Pro Highspeed, arrives with it.

A note on the calendar: the desk's event date is October 23, and Unite.AI published its report that day. The article text, however, says Moonshot introduced the model on October 24. The cited report does not explain the gap, so I will not pretend to.

What is actually open

Here is the first footnote. Moonshot's blog says full model weights "will be released by Nov 4, 2026." The model is available today through Kimi's own products and API. The weights are a promise with a date on it.

The announcement material available to the desk gives no license. It does not say whether Highspeed gets open weights too, and it does not state hardware requirements for running the thing yourself. Wallstreetcn reports 2.8 trillion total parameters in a Mixture of Experts design, activating 104 billion per token. Moonshot's own blog sticks to "3T-class." Those figures are compatible, but only one of them is a number.

According to Moonshot, K3.1 Pro is a post-trained version of Kimi K3 with substantially larger post-training compute. The company says it reworked its reinforcement learning stack, from long-context inference to sandbox infrastructure, so joint RL across hundreds of downstream domains pays off predictably.

The numbers, sorted by who graded the exam

Everything below is Moonshot's reporting of its own results. Sorted by who wrote the test:

• Terminal-Bench 4.0: 65.7 percent for K3.1 Pro, ahead of GPT-6 Astra at 58.2 and narrowly behind Claude Opus 5.5 at 66.4.
• BenchCAD (with tools): 92.7 percent, behind Opus 5.5 at 96.2 and GPT-6 Astra at 95.9.
• OfficeQA Pro v2: 64.4, which the company calls state of the art. The announcement gives no comparison scores for it.
• Kimi Code Bench 3.0 Long-horizon: 61.2 percent against K3's 55.95. This is Moonshot's internal benchmark.
• Kimi Hallu Bench Hard: claim-level hallucination fell from 23.4 percent for K3 to 8.3 percent. Responses with at least one factual error ran at 42.4 percent of K3's rate. Also internal.

The pattern is familiar. On Terminal-Bench, K3.1 Pro beats GPT-6 Astra but trails Claude Opus 5.5. On BenchCAD, it trails both, which Moonshot calls "approaching." That is respectable, not an across-the-board win over both rivals.

The finance results come with their own discrepancy. Unite.AI and Moonshot's blog say K3.1 Pro improved on K3 by 7.2 points on Financial Modeling and 11.6 points on Financial Research, and scored about 4.2 times K3 on Business Operation. Wallstreetcn reports the first two as 14 percent and 20 percent. Points and percent are different units, and the desk cannot reconcile them from what has been published.

The most entertaining result is a human walking simulator. In an internal test, K3.1 Pro built one on the open-source human-atlas model (296 skeletal components, 402 muscle components). Over 167 missions and 420 independent reviews, it got the simulated human to walk for 11.2 seconds. Moonshot says that when reward shaping stalled, the model probed the simulator's dynamics and fixed 36 errors in shoulder-muscle moment-arm calculations. Eleven seconds is not a marathon, but a model that debugs the physics instead of fiddling with the reward is the correct instinct.

Cyber data, removed on purpose

Moonshot says it removed cyber-related data from both mid-training and RL before the open-weight release. On its internal xdev-bench-hard, a checkpoint without those measures solved 90.5 percent of problems at pass@3. The final model reaches 81.0 percent.

The company states two principles: not to release open weights that extend the absolute frontier of cyber capability, and to progressively raise the cyber capability available in open weights for defensive auditing. A 9.5-point drop is a real, measurable cost. Whether it is enough is another question: the announcement supplies one internal benchmark, run by the party doing the removing.

Price and access

On OpenRouter, K3.1 Pro lists at $2.00 per million prompt tokens and $10.00 per million completion tokens, with a 1,048,576-token context. An X post from @kimmonismus calculates that this is 50 percent below Claude Opus 5.5 ($4/$20) and 80 percent below GPT-6 Astra ($10/$50). That is the poster's arithmetic, though the division checks out.

Moonshot's own standard API prices, as reported by Wallstreetcn, are 1.3 yuan per million tokens for cache hits, 13 yuan for cache misses and 68 yuan for output. Highspeed costs double and runs two to three times faster, with identical capabilities, according to the announcement.

The model takes image and video input and supports reasoning_effort settings of low, high or max, with max the default. It supports thinking mode only: Kimi Code documentation says requests with thinking disabled get an HTTP 400 error. Enterprise access is off by default until an administrator turns it on.

What to watch

Three things. First, November 4: do the weights show up, under what license, and can anyone outside a data center run 2.8 trillion parameters? Second, whether independent evaluators reproduce the Terminal-Bench and BenchCAD numbers. Third, whether the hallucination drop survives a test Moonshot did not write.

Open weights are a good thing to promise. Show me the eval, not the vibe, and for now also show me the download.