New to this? Read it in simple words
- Moonshot, a Chinese AI company, released a new model called Kimi K3.1 Pro.
- An outside tester ranks it fourth of 215 models, with 55 points.
- Moonshot says anyone will be able to download it by November 4, 2026.
- Moonshot says it cut hacking material from training, but the model still hacks well.
- Open weights
- A model’s learned numbers published for anyone to download, run and change on their own computers.
- Token
- A word or part of a word. AI models read and write text in tokens.
- Artificial Analysis
- An independent company that runs the same set of tests on AI models and compares them.
Moonshot AI, a Chinese company based in Beijing, makes the Kimi chatbot and the Kimi models. Its earlier model, K3, came out in July. On Friday, October 23, it released an upgrade, Kimi K3.1 Pro.
What Moonshot released
Kimi K3.1 Pro works with text and images, and can read up to 1 million tokens at once. Moonshot calls it its most capable “3T-class” open-weight model. Artificial Analysis lists it with 2.8 trillion parameters.
A second version, K3.1 Pro Highspeed, runs two to three times faster, Moonshot says. Both are available now on Kimi.ai, in the Kimi apps, in Kimi Work and Kimi Code, and through the Kimi API.
Moonshot says it will release the full model weights by November 4, 2026. Until then, nobody outside the company can download the model, and Artificial Analysis lists it as proprietary.
Moonshot’s claims beside Artificial Analysis’s test of October 23.
Moonshot’s own test results
On Terminal-Bench 4.0, a test of coding agents, Moonshot says K3.1 Pro scored 65.7%. That puts it ahead of OpenAI’s GPT-6 Astra at 58.2%, and just behind Claude Opus 5.5 at 66.4%.
Moonshot also says the model makes far fewer false claims on its own stress test: 8.3% of answers, down from 23.4% for K3. These are Moonshot’s own figures, and nobody outside has checked them, TNW notes.
What the outside test found
Artificial Analysis runs its own set of 10 tests on every new model. It gives K3.1 Pro 55 points on its Intelligence Index and ranks it fourth of the 215 models it compares it with.
The tester found the model slower than average, at 73.5 words or word parts (tokens) per second. It is also very wordy: it wrote 150 million tokens during the tests, against a typical 87 million.
Through Moonshot’s API, it costs $2 per million input tokens and $10 per million output tokens, Artificial Analysis says. One run of its index cost $2.11 per task on average.
Sources2
Less hacking skill, by design
Weights that anyone can download can’t be called back. Moonshot says it therefore removed material about cyberattacks from the training data before the open release.
On Moonshot’s in-house test of attacks on real software flaws, a version trained without these cuts solved 90.5% of the tasks. The released model solved 81%, and K3 solved 33.3%.
So by Moonshot’s own numbers, the cut is small and the model is still much stronger at this than its predecessor. DeepSeek’s V4.1 Flash scored 85.7% on the same test, Moonshot says.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · October 23, 2026Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence Moonshot AI (Kimi)
- 2Research · October 23, 2026Kimi K3.1 Pro: Intelligence, Performance & Price Analysis Artificial Analysis
- 3
- 4Research · October 23, 2026Moonshot AI Releases Kimi K3.1 Pro, Beats Opus 5.5, GPT Astra 6 On Some Benchmarks OfficeChai
The coding, false-claim and hacking scores are Moonshot’s own. The index score, speed and price come from Artificial Analysis. The open weights are still a promise.
A major model launch ranked fourth of 215 by an independent tester, tempered by unreleased weights and still-strong hacking skills. How we rate the mood