Moonshot ships its first Pro model
Kimi K3.1 Pro：月之暗面端出第一个Pro
B · Useful, limited Jiqizhixin launch coverage of Moonshot's Kimi K3.1 Pro that combines independent Artificial Analysis scores with vendor benchmarks and hands-on tests; informative, but mostly a relay of launch material written in an enthusiastic register.
Summary based on the full text·Original: about 9 minutes to read
In short
A Jiqizhixin article (24 October 2026) on Kimi K3.1 Pro, the first "Pro" model from Moonshot AI and the first of its K3.1 series, at around 3 trillion parameters with a 1 million token context. It is large-scale reinforcement-learning post-training on K3, the 2.8-trillion-parameter open-weight model released in July. Independent tester Artificial Analysis ranks it fourth on its intelligence index, behind Claude Opus 5.5, and top on its Terminal-Bench 4.0 run. Full weights are promised by 4 November.
Key numbers
- 55 vs 58Artificial Analysis intelligence index for K3.1 Pro (4th worldwide) versus Claude Opus 5.5 (1st)
- 67% vs 13%Artificial Analysis Terminal-Bench 4.0 score for K3.1 Pro versus K3, the highest of all models tested
- 12 to 3 pointsgap between the best open-weight and closed models on the index once K3.1 Pro weights ship, per Artificial Analysis
- 23.4% to 8.3%per-claim factual error rate on Moonshot's internal Kimi Hallu Bench Hard stress test, K3 versus K3.1 Pro
- 1.3 / 13 / 68 yuanAPI price per million tokens: cached input, uncached input, output (about a third below K3)
As reported by the source.
Chinese·Web page·opens jiqizhixin.com
What it reports
- Artificial Analysis: intelligence index 55 (up 11 from K3), fourth behind Claude Opus 5.5 (58), Claude Sonnet 5.5 and Muse Wonder Full (56). Once weights are out it would be the top open-weight model, closing the open-closed gap from 12 to 3 points.
- Largest gains are in agent work: GDPval-AA Elo 1846 (+312, third); Terminal-Bench 4.0 67% versus 13% for K3, highest of all models; AutomationBench-AA 58% to 73%. Humanity's Last Exam rose more modestly, 47% to 52%.
- Vendor figures: Terminal-Bench 4.0 65.7% versus GPT-6 Astra 58.2% and Opus 5.5 66.4%; best score on 16 of 47 listed benchmarks (3 internal); SWE Marathon v1.1 51.7. Still behind closed flagships on HLE-Diamond and ProgramBench.
- 3D and front-end work: BenchCAD 92.7% with tools (Opus 5.5 96.2%); scripts Blender and FreeCAD directly rather than through connectors.
- Internal case: building a human walking simulator, the model found 36 errors in shoulder-muscle moment-arm calculations by probing the simulator's dynamics.
- Finance: OfficeQA Pro v2 64.4 versus 62.2 for Opus 5.5. On Moonshot's internal hallucination stress test the per-claim error rate fell from 23.4% to 8.3% (Opus 5.5 15.0%, GPT-6 Astra 5.5%), which Moonshot says does not reflect everyday error rates.
- Price per million tokens: 1.3 yuan cached input, 13 yuan uncached input, 68 yuan output, about a third below K3 and roughly a fifth of GPT-6 Astra's list price. A high-speed version costs twice as much.
- Enterprise options: zero data retention for API accounts, an enterprise edition off by default, hosted agents, and a partner programme for forward-deployed engineers.
How they got it
The article combines Artificial Analysis's published results, Moonshot's official blog and benchmark table, and the editors' own one-shot tests (an SVG animation, a Three.js page, a Godot game, Blender models). The independent and vendor figures are distinguished in the text.
Limits
- Most benchmarks are vendor-reported, including three internal ones and the hallucination test.
- Hands-on tests are qualitative showcases chosen by the editors.
- Weights were not yet released at publication, so open-weight claims are prospective.
Why it matters for companies
If the weights ship as promised, firms gain a near-frontier model they can run in their own data centres, fine-tune and keep data on-site, at a list price far below closed flagships. The gains concentrate in agent and terminal tasks, which is where most enterprise automation effort now sits. The vendor's own note that its hallucination stress test is not a real-world error rate is a useful reminder to evaluate on your own documents before using any model on finance work.
Key passages in translation
Chinese original
K3.1 Pro 拿下 55 分，较 K3 提升 11 分，位列全球第四
K3.1 Pro scored 55, 11 points above K3, ranking fourth worldwide.
Chinese original
在 Artificial Analysis 自己的 Terminal-Bench 4.0 测试中， 它以 67% 的成绩拿下所有模型中的最高分 ，而 K3 只有 13%
In Artificial Analysis's own Terminal-Bench 4.0 test it scored 67%, the highest of all models, while K3 scored only 13%.
Chinese original
官方也特别注明，这是刻意刁难模型的压力测试，不代表日常使用中的整体错误率。
The company also notes that this is a deliberately adversarial stress test and does not represent the overall error rate in everyday use.
Summary based on the full text. Read the original for the full argument and data.Read the original on jiqizhixin.com