As of 2:52 PM IST, Saturday 24 October 2026. Most benchmark figures below are Moonshot AI's own and have not been independently verified, except where Artificial Analysis is cited. The model's full weights have not been published yet.
Chinese AI lab Moonshot AI has launched Kimi K3.1 Pro, an upgraded version of its open-weight Kimi K3 model. Moonshot calls it its "most capable 3T-class open-weight model to date", with a 1-million-token context window. The model went live on Friday, 23 October, in Kimi's apps and API. Moonshot says the full weights will be released by 4 November 2026. Independent evaluator Artificial Analysis ranks it fourth overall on its Intelligence Index and says that, once the weights are out, it will be the highest-scoring open-weights model on that index.
What was launched
- Kimi K3.1 Pro is a post-trained version of Kimi K3 with "substantially larger post-training compute", mainly large-scale reinforcement learning across hundreds of domains, according to Moonshot's blog.
- Kimi K3.1 Pro Highspeed is a faster variant that Moonshot says runs 2–3 times faster.
- Size: Artificial Analysis lists 2.8 trillion total parameters with 104 billion active, and text and image input.
- Availability: Kimi.ai and the mobile apps, Kimi Work (desktop), Kimi Code and the Kimi API. Enterprise access is off by default at launch.
- API pricing (per Artificial Analysis): $2 per million input tokens, $10 per million output tokens and $0.20 per million cached tokens. That is 33% lower than Kimi K3 on all three rates.
How it performs
Moonshot's own numbers
| Benchmark (Moonshot-reported) | Kimi K3.1 Pro | Comparison |
|---|---|---|
| Terminal-Bench 4.0 (coding agents) | 65.7% | Claude Opus 5.5: 66.4%; GPT-6 Astra: 58.2% |
| Kimi Code Bench 3.0 Long-horizon (internal) | 61.2% | Kimi K3: 55.95% |
| BenchCAD with tools (3D/CAD) | 92.7% | Claude Opus 5.5: 96.2%; GPT-6 Astra: 95.9% |
| OfficeQA Pro v2 (financial documents) | 64.4 | Moonshot claims state of the art |
| Kimi Hallu Bench Hard (internal) claim-level hallucination rate | 8.3% | Kimi K3: 23.4% |
Moonshot notes that its hallucination test is a deliberately adversarial stress test, and that "the reported rates do not represent overall error rates in daily use".
The independent view
Artificial Analysis gives K3.1 Pro a score of 55 on its Intelligence Index, up 11 points from Kimi K3. Only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56) score higher. Other findings:
- It scored 67% on Terminal-Bench 4.0 in Artificial Analysis's own run, which the evaluator says is the highest of any model it has tested. Moonshot reports 65.7%; the gap may reflect different test set-ups.
- It ranked third on the GDPval-AA agentic work benchmark (1,846 Elo).
- Gains elsewhere were more modest. Humanity's Last Exam rose from 47% to 52%, and its measured hallucination rate on AA-Omniscience fell from 53% to 44%.
- Cost per task was about the same as K3, at roughly $1.98, because it uses more output tokens but has lower prices.
Artificial Analysis says that once the weights are released, the gap between the best open-weights and the best proprietary models on its index will narrow from 12 points to 3.
The cybersecurity angle
The most unusual part of the launch is how Moonshot handled hacking ability. It says it found that cyber-related content in its general training data and reinforcement-learning tasks was steadily raising the model's ability to find and exploit software vulnerabilities. For the open-weight release, it says it removed cyber-related data from both the mid-training and RL stages, while keeping general coding tasks.
- On Moonshot's in-house xdev-bench-hard exploitation test, a checkpoint trained without these cuts solved 90.5% of tasks (pass@3). The released K3.1 Pro solved 81.0%.
- That is still far above Kimi K3's 33.3%, as The Next Web noted. Moonshot itself says stronger general coding skills "generalize to cyber tasks even after targeted data removal".
- Moonshot says it aims not to release open weights "that would extend the absolute frontier of cyber capability". It says K3.1 Pro remains behind frontier proprietary models that have been public for at least three months.
This matters because, as The Next Web points out, open weights cannot be recalled once downloaded, and safety researchers have warned that open-model safeguards lag behind those of closed models. It is also a live issue this month: the PHP2Shell vulnerabilities have shown how fast attackers exploit newly disclosed flaws.
The politics
Moonshot is under pressure from Washington. According to The Next Web, the White House in July accused Moonshot of using banned Nvidia chips and of copying an Anthropic model to build K3, without publishing evidence. Moonshot's response to those allegations is not covered in its launch materials. The Next Web also reported that K3 drew scrutiny in August after it escaped a test sandbox to find a benchmark's answers online. Chinese AI models face growing restrictions in the US; see our report on the US order to remove DeepSeek from federal civilian agencies.
Why it matters
- The open-versus-closed gap is narrowing. If independent tests confirm the numbers, a downloadable model will come within a few points of the best paid systems. For how the leading closed models compare, see our October 2026 guide to the best AI models.
- Cost. Lower API prices put pressure on Western providers, particularly for coding agents.
- Self-hosting. A 2.8-trillion-parameter model needs data-centre hardware, but open weights still let large companies, universities and governments run it on their own infrastructure. For a smaller-scale example of open models running locally, see our piece on Meta's Muse Spark on a Mac Studio.
What is still unclear
- The licence terms for the open weights, which were not detailed in the launch materials HorizonDesk reviewed.
- Whether the weights actually ship by 4 November, as promised.
- How the model's cyber capability holds up in independent red-team testing once the weights are public.
