AnswersPublished27 sources
What Kimi K3.1 Pro Claims—and What Its Launch Data Doesn’t Prove
Launched October 23, 2026, Kimi K3.1 Pro is Moonshot’s first K3.1 model, with a 1 million token context window and a faster Highspeed version. Moonshot reports 65.7% on Terminal Bench 4.0, between GPT 6 Astra at 58.2% and Claude Opus 5.5 at 66.4%; treat the comparison as a vendor reported result, not an independent...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Launched October 23, 2026, Kimi K3.1 Pro is Moonshot’s first K3.1 model, with a 1 million token context window and a faster Highspeed version.
- Moonshot reports 65.7% on Terminal Bench 4.0, between GPT 6 Astra at 58.2% and Claude Opus 5.5 at 66.4%; treat the comparison as a vendor reported result, not an independent ranking.
- The model targets coding, financial research and 3D work, but the license for the planned weight release and the real world significance of its internal factuality tests remain unclear.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What are the key details of Beijing-based, $50-billion-valued Moonshot AI’s Friday, October 23 launch of Kimi K3.1 Pro—the first model in it. Article summary: Moonshot launched Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed on October 23, 2026, positioning Pro as the first K3.1-series model and its most capable “3T-class open-weight” model, with a 1-million-token context window.[2]. Topic tags: general, general web, user generated, news. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts w
Moonshot AI launched Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed on October 23, 2026. The company describes Pro as the first model in its K3.1 series and its most capable “3T-class open-weight” model, with a 1-million-token context window. Bloomberg reported that the upgrade is aimed at financial research, coding and 3D modeling, including work related to game design.17
18
The headline specifications and benchmark numbers make K3.1 Pro worth evaluating. They do not, by themselves, establish that it outperforms competitors across everyday tasks: several widely circulated results are Moonshot-reported, and the full weights had only been promised for a later date at launch.8
10
What Moonshot says changed
K3.1 Pro is described as a post-trained version of Kimi K3, with additional post-training investment, rather than an entirely new base model.21
25 Launch materials describe a mixture-of-experts architecture with 2.8 trillion total parameters and 104 billion active parameters. In other words, the “3T-class” label refers to the model’s overall scale, not the number of parameters active at once.
23
30
Moonshot is pitching the model for coding, financial research and 3D creation. Its announcement names tools and environments including Blender, Unreal Engine, Godot and CAD.17
30 That breadth is a company-stated product focus; it is not a guarantee that the model will perform equally well across all those workflows.
How to read the benchmark claims
Moonshot reports a 65.7% score for K3.1 Pro on Terminal-Bench 4.0. The comparison figures reported alongside it are 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5—placing K3.1 Pro above the former and 0.7 percentage points below the latter in that comparison.2
6
For 3D work, Moonshot reports 92.7% on BenchCAD with tools. The test is described as assessing 3D reconstruction from multiple views through CAD-code generation.29 Moonshot also reports a score of 64.4 on OfficeQA Pro v2, a financial-document evaluation.
10
46
These scores are useful clues about the tasks Moonshot chose to highlight, not a complete measure of general capability. The Terminal-Bench comparison is reported through launch coverage, while the 3D and finance figures are attributed to Moonshot. Readers choosing a model for a real workflow should test the same tasks with the same tools, settings and success criteria.
Factuality claims need context
Moonshot’s materials and launch coverage cite improvements in factual accuracy, but the reported figures do not all use the same description of the evaluation. One account of an internal hard factuality test reports that claim-level hallucinations fell from 23.4% for Kimi K3 to 8.3% for K3.1 Pro. The Kimi API documentation describes a separate evaluation using de-identified conversations in which users had flagged factual errors, and reports a reduction from 23.35% to 6.85%.18
21
Those numbers should not be merged or presented as an everyday error rate. They refer to evaluations with specific prompts, samples and scoring methods; claim-level errors also differ from counting how many entire responses contain at least one mistake. The figures are evidence of reported test results, not a guarantee of accuracy in a user’s own work.
Moonshot also says it removed cyber-related training data to limit offensive cyber capability in the planned open-weight model.13 That describes a mitigation effort, not proof that offensive capabilities have been eliminated or that the intervention is effective in every setting.
Highspeed, access and the weight release
Moonshot describes Kimi K3.1 Pro Highspeed as offering roughly two to three times the speed of Pro while maintaining equivalent intelligence.18
20 That speed claim is not the same as an independent comparison of quality or latency across real workloads.
Kimi Code documentation says the K3.1 Pro models operate in thinking mode, with low, high and max settings; max is the default.33 Launch coverage lists availability through Kimi.ai, Kimi Work, Kimi Code and the Kimi API.
10
30 For Kimi Code, reporting says standard Pro requires Plus or higher and Highspeed requires Pro or higher; the full 1-million-token context is listed for Pro-tier access or above.
34
36
Moonshot promised to release the full weights by November 4. That was a planned date, not confirmation that downloadable weights were available at launch.8
10 The available launch information does not establish the final license. “Open-weight” alone does not tell users whether commercial use, redistribution or modification will be permitted; those terms depend on the license actually released.
The practical takeaway
Kimi K3.1 Pro’s launch case rests on a large context window, claimed gains in coding, finance and 3D tasks, and benchmark results that look competitive in selected comparisons. The key caveat is that headline scores and internal factuality results need to be tested in context, while the weight release and its license were still outstanding at launch. Independent, matched evaluations and the final license will matter more to many users than the “3T-class” label alone.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What Kimi K3.1 Pro Claims—and What Its Launch Data Doesn’t Prove"?
Launched October 23, 2026, Kimi K3.1 Pro is Moonshot’s first K3.1 model, with a 1 million token context window and a faster Highspeed version.
What are the key points to validate first?
Launched October 23, 2026, Kimi K3.1 Pro is Moonshot’s first K3.1 model, with a 1 million token context window and a faster Highspeed version. Moonshot reports 65.7% on Terminal Bench 4.0, between GPT 6 Astra at 58.2% and Claude Opus 5.5 at 66.4%; treat the comparison as a vendor reported result, not an independent ranking.
What should I do next in practice?
The model targets coding, financial research and 3D work, but the license for the planned weight release and the real world significance of its internal factuality tests remain unclear.