Futureweb

Artificial Intelligence

Moonshot's Kimi K3.1 Pro is on the company's API now. The post says full weights come by November 4. The benchmark figures below are the company's, including an internal cyber comparison it says is lower by design.

Kimi K3.1 Pro is on Moonshot's API now. The company says full weights come by November 4, so they are not public today. The scores are the company's.

The short version

No. Moonshot says the model and a faster variant are available on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, and that full weights will be released by November 4, 2026. The post calls it a 3T-class open-weight model with a one-million-token context. Scores on the post are the company's. Its internal cyber comparison, 81.0 percent against 90.5 percent without the filter, is the company's and is not independently verified here.

Hosted now, weights by November 4

Moonshot's post introduces Kimi K3.1 Pro as the first model in the K3.1 series and its most capable 3T-class open-weight model so far, with a one-million-token context. It also introduces Kimi K3.1 Pro Highspeed, which it says delivers two to three times the speed of K3.1 Pro. Both, the post says, are available on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Full model weights will be released by November 4, 2026. On October 24 that file is a date, not a download. Simon Willison's October 23 note and Latent Space's October 24 note are outside accounts of the same post. Latent Space's headline states 104 billion active parameters and 2.8 trillion total. The company post's phrase is 3T-class. This page will not pretend those are one sentence.the Geo Key beta is a different October 23 post.

The scores are Moonshot's

On Terminal-Bench 4.0 the post reports 65.7 percent, against GPT-6 Astra at 58.2 percent and Claude Opus 5.5 at 66.4 percent. On Kimi Code Bench 3.0, long horizon, it reports 61.2 percent against K3 at 55.95 percent. Those comparisons are the company's. A lab that later scores the weights, once the weights exist, can disagree. Until November 4, an outside lab is scoring an API Moonshot operates, or it is not scoring the model at all. This page did not run the bench. Open-weight, in the post, is a promise with a deadline. It is not a weight file a reader can hash today.Dessa prints neither a score nor a size.

The cyber paragraph is a limit, not a guide

The post says cyber-related material was removed from mid-training and from the reinforcement-learning mix for this open-weight release, while general coding tasks were kept. On a test it calls xdev-bench-hard, it says a checkpoint without that removal solves 90.5 percent of problems at pass at 3, and the final K3.1 Pro reaches 81.0 percent. It says the point of the comparison is that removal reduces cyber capability relative to the same pipeline, and that the model is still behind frontier proprietary models that have been public for at least three months. It says the aim is not to release open weights that would extend the absolute frontier of cyber capability. This page does not describe the problems, the data, or the filter. The figures are the company's, and they are not independently verified here.Nemotron's 4GB page is a different disclosure.

What the outside notes are for

Willison and Latent Space are how the post left Moonshot's site and reached a wire. They are not a second training run. Where a headline prints 2.8 trillion and 104 billion active, that is the headline. Where the post prints 3T-class and a November 4 weight date, that is the post. A reader who needs one specification should notice the post did not use the headline's pair. Highspeed's two-to-three-times claim is the company's, about its own API, not about the weights that are not out.a context window is the bound the post is using.

Who is in scope

People using the named Kimi products and the API today, under whatever terms those products have. People waiting on weights, who have a deadline and not a file. People who wanted a cyber how-to, who will not find one here or, if the company's paragraph is true, in the training mix it says it removed. Enterprise defaults and pricing, where a card prints them, are that card's. This brief does not copy a rate the post's text, as used here, does not make the subject.

A lower internal score is not a safety proof

81.0 against 90.5 says the company measured a drop on its own test after a removal. It does not say the model cannot do the work, and the post says general coding still generalizes into that work. It does not say an outside auditor agreed. A designed reduction on a private benchmark is a design claim. Treating it as a clearance would invent the audit.

The date that matters

November 4, 2026 is the company's deadline for weights. October 23 and October 24 are the note and the wire. The model on the API is a present-tense product. The open-weight label is future tense until the file exists. China's export-control drafts, when a wire mentions them, are background a reporter adds. They are not a term printed as a condition of this model card in the sentences this brief relies on.on-device weights are a different promise.

What this page will not merge

It will not merge Latent Space's 2.8 trillion with the post's 3T-class. It will not merge another lab's Terminal-Bench score, if one appears, with 65.7. It will not describe the internal items. It will not call the weights released. The record is the post.

The limit

Hosted now. Weights by November 4. Company scores labeled. Cyber figures labeled and not a guide. The file is not on the desk. The short answer this page will stand on is this: No. Moonshot says the model and a faster variant are available on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, and that full weights will be released by November 4, 2026. The post calls it a 3T-class open-weight model with a one-million-token context. Scores on the post are the company's. Its internal cyber comparison, 81.0 percent against 90.5 percent without the filter, is the company's and is not independently verified here.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. Moonshot
  2. Simon Willison
  3. Latent Space

Questions

Can a reader download the weights today?

The post says full weights will be released by November 4, 2026. Until that release, the model the post describes is the hosted one. A promise of a file is not the file.

What does the Latent Space headline add?

Its headline states 104 billion active parameters and 2.8 trillion total. The company post says 3T-class. This page does not collapse those into one specification.

Whose Terminal-Bench number is 65.7 percent?

Moonshot's, for Terminal-Bench 4.0, set beside GPT-6 Astra at 58.2 percent and Claude Opus 5.5 at 66.4 percent. Those are the post's figures. Another lab's score of the same test would be that lab's.

What does the company say it removed?

Cyber-related data from mid-training and cyber-related tasks from the reinforcement-learning mix, for the open-weight release. The post says the goal is not to release open weights that extend the absolute frontier of cyber capability. This page does not describe the tasks.

Is 81 percent a public exam?

It is the company's internal comparison, xdev-bench-hard, pass at 3: 81.0 percent for the released checkpoint and 90.5 percent for a checkpoint without the removal. Not independently verified on this page.