据该公司称，除了 K3.1 Pro 之外，Moonshot 还推出了一个名为 K3.1 Pro Highspeed 的变体，其运行速度是前者的两到三倍。目前，这两个模型均可通过 Kimi.ai、Kimi Work、Kimi Code 和 Kimi API 使用。
K3.1 Pro 最显著的改进体现在代理任务上。在 GDPval-AA 基准测试中，根据 Artificial Analysis 的数据，其达到 1,846 Elo，比 K3 高出 312 分。这使其仅次于 Claude Opus 5.5（1,866）和 Muse Wonder Full（1,857）。在 Terminal-Bench 4.0 上，Moonshot 报告得分为 65.7%，领先于 GPT-6 Astra（58.2%），并略低于 Claude Opus 5.5（66.4%）。而在同一测试中，K3 的得分仅为 13%。
幻觉率也大幅下降。据该公司称，在 Moonshot 内部的 Kimi Hallu Bench Hard 测试中，声明级别的幻觉从 K3 的 23.4% 降至 8.3%。Moonshot 表示这是一项压力测试，而非衡量一般错误率的指标。GPT-6 Astra 在同一基准测试中的得分为 5.5%，因此它仍具有优势。
在内部金融基准测试中，K3.1 Pro 在金融建模方面提升了 7.2 分，在金融研究方面提升了 11.6 分。对于通用办公任务，据 Moonshot 称，该模型在 OfficeQA Pro v2 上超越了所有受测模型。
据 Moonshot 称，K3.1 Pro 可以直接处理 Unity、Unreal Engine 和 Godot 等游戏引擎，既能进行快速修复，也能完成完整的项目构建。在 BenchCAD 基准测试中，使用工具时其得分为 92.7%，接近 Claude Opus 5.5（96.2%）和 GPT-6 Astra（95.9%）。不使用工具时，其得分为 72.8%，几乎与 Claude Opus 5.5（73.0%）持平。
Moonshot 表示，该模型在机械零件的参数化建模、建筑模型和可配置产品方面表现最强。它可以编写并运行 Blender 和 CAD 软件的脚本，无需 MCP 连接器。该公司还在 Kimi.com 上推出了一项名为“3D Make”的功能，允许用户预览、编辑和导出 3D 模型。
K3.1 Pro 保持了与 K3 相同的尺寸，总参数量为 2.8 万亿，活跃参数量为 1040 亿。上下文窗口涵盖一百万个 token，且模型同时支持文本和图片。据 Artificial Analysis 称，它位于帕累托前沿的顶端。没有参数量相当的模型在智能指数上得分更高。
Kimi K3.1 Pro 以略低的单任务成本（$1.98 vs. $2.00）提供了远高于 Kimi K3 的性能，使其位于 Artificial Analysis 的帕累托线上。| 图片来源：AA
Token 价格下降了 33%。输入成本为每百万 token 2 美元（从 3 美元下降），输出成本为 10 美元（从 15 美元下降），缓存命中成本为 0.20 美元（从 0.30 美元下降）。每个智能指数任务的成本为 1.98 美元，与 K3（2.00 美元）几乎相同，因为该模型生成的输出 token 多了 14%。
对于开源权重的发布，Moonshot 从两个训练阶段中移除了与网络相关的数据。在其内部的 xdev-bench-hard 测试中（该测试重现了 Linux 内核、Node.js 和 JavaScriptCore 等软件中最近披露的漏洞），未经清理的版本解决了 90.5% 的任务。数据清理后，K3.1 Pro 解决了 81.0%。DeepSeek-V4.1-Flash 在同一测试中达到 85.7%，而 Kimi K3 的得分为 33.3%。
查看原文
Moonshot AI's new reasoning model Kimi K3.1 Pro lands at number four on the Artificial Analysis Intelligence Index, just 3 points behind the top proprietary models.
Chinese AI company Moonshot AI has released Kimi K3.1 Pro, a post-trained version of its predecessor Kimi K3 with significantly more post-training compute. The model scores 55 points on the Artificial Analysis Intelligence Index, an 11-point jump over K3 that lands it in fourth place.
Only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder Full (56) rank higher, according to Artificial Analysis. If Moonshot releases the model weights as planned by November 4, 2026, K3.1 Pro would become the highest-scoring open-weights model available. That would shrink the gap between open and proprietary models from 12 to just 3 points, assuming nothing else changes by then.
From midfield to the top four: Kimi K3.1 Pro gains a full 11 points over Kimi K3 on the Artificial Analysis Intelligence Index. | Image: AA
Alongside K3.1 Pro, Moonshot is shipping a variant called K3.1 Pro Highspeed that runs two to three times faster, according to the company. Both are available now through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Item
Kimi-K3.1-Pro
Kimi-K3.1-Pro-Highspeed
Input
$2.00 / MTok
$4.00 / MTok
Output
$10.00 / MTok
$20.00 / MTok
Cache Input
$0.20 / MTok
$0.20 / MTok
Agentic tasks see the biggest gains
K3.1 Pro's strongest improvements show up in agentic tasks. On the GDPval-AA benchmark, it reaches 1,846 Elo according to Artificial Analysis, 312 points above K3. That puts it just behind Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857). On Terminal-Bench 4.0, Moonshot reports 65.7 percent, ahead of GPT-6 Astra (58.2 percent) and barely behind Claude Opus 5.5 (66.4 percent). K3 managed only 13 percent on the same test.
Hallucination rates dropped sharply too. On Moonshot's internal Kimi Hallu Bench Hard, claim-level hallucinations fell from 23.4 percent with K3 to 8.3 percent, according to the company. Moonshot says this is a stress test, not a measure of general error rates. GPT-6 Astra hits 5.5 percent on the same benchmark, so it still has an edge.
On internal finance benchmarks, K3.1 Pro gained 7.2 points on Financial Modeling and 11.6 points on Financial Research. For general office tasks, the model tops all tested models on OfficeQA Pro v2, according to Moonshot.
Benchmark
Kimi K3.1 Pro
Claude Opus 5.5
GPT-6 Astra
Claude Fable 5.1
DeepSeek-V4.1-Flash
GLM-5.3
Terminal-Bench 4.0
65.7
66.4
58.2
57.9
42.3
41.8
SWE Marathon v1.1
51.7
46.3
42.5
45.6
40.0
42.5
DeepSWE v1.1
75.5
74.2
74.1
67.4
77.0
66.9
BenchCAD (with tools)
92.7
96.2
95.9
92.6
81.6
--
HLE-Full (with tools)
70.0
67.7
57.2
65.6
58.8
62.5
BrowseComp
92.1
90.8
91.5
--
--
--
OfficeQA Pro v2
64.4
62.2
54.4
54.5
40.4
31.1
Automation Bench-AA
73.3
69.5
68.5
59.4
66.6
62.2
GDPval-AA v2.1 (Elo)
1,846
1,866
1,573
1,754
1,746
1,650
GPQA-Diamond
94.2
93.9
96.1
93.7
92.2
91.7
ExploitBench
74.7
91.0
100
83.0
--
54.4
Moonshot's Kimi K3.1 Pro benchmark table. All results with reasoning effort set to "max."
3D modeling and coding
K3.1 Pro works directly with game engines like Unity, Unreal Engine, and Godot, handling both quick fixes and full project builds, according to Moonshot. On the BenchCAD benchmark, it scores 92.7 percent with tools, close to Claude Opus 5.5 (96.2 percent) and GPT-6 Astra (95.9 percent). Without tools, it hits 72.8 percent, nearly matching Claude Opus 5.5 (73.0 percent).
Moonshot says the model is strongest at parametric modeling of mechanical parts, architectural models, and configurable products. It can write and run scripts for Blender and CAD software without MCP connectors. The company also launched a feature called "3D Make" on Kimi.com that lets users preview, edit, and export 3D models.
Same size, lower prices
K3.1 Pro keeps the same dimensions as K3, with 2.8 trillion total parameters and 104 billion active ones. The context window spans one million tokens, and the model handles both text and images. According to Artificial Analysis, it sits at the top of the Pareto frontier. No model with a comparable parameter count scores higher on the Intelligence Index.
Kimi K3.1 Pro delivers much higher performance than Kimi K3 at slightly lower cost per task ($1.98 vs. $2.00), placing it on the Artificial Analysis Pareto line. | Image: AA
Token prices drop by 33 percent. Input costs $2 per million tokens (down from $3), output costs $10 (down from $15), and cache hits cost $0.20 (down from $0.30). Each Intelligence Index task costs $1.98, almost identical to K3 ($2.00), since the model generates 14 percent more output tokens.
Cybersecurity and deliberate limits on offensive capabilities
Moonshot devotes an entire section of its technical blog post to cybersecurity. The company says it observed that vulnerability research capabilities kept growing as post-training scaled up, because cyber-related content was embedded in the general mid-training corpus and related tasks were part of the RL environment mix.
For the open-weights release, Moonshot stripped cyber-related data from both training phases. On its internal xdev-bench-hard, which reproduces recently disclosed vulnerabilities in software like the Linux kernel, Node.js, and JavaScriptCore, the uncleaned version solves 90.5 percent of tasks. After data cleaning, K3.1 Pro solves 81.0 percent. DeepSeek-V4.1-Flash reaches 85.7 percent on the same test, while Kimi K3 scored 33.3 percent.
Moonshot says no open-weights model should push the absolute frontier of offensive cyber capabilities. At the same time, the company wants to gradually raise what open models can do, since they also serve defensive purposes like auditing legacy systems. Anthropic previously showed that even GLM-5.3, which scores 45 on the AA Index, can write exploits at near-Claude Mythos levels.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
Studio