Moonshot发布Kimi K3.1 Pro,Claude Duet预览
🤖 AI Agents Weekly: Claude Duet, Kimi K3.1 Pro, MiniMax M3.1 Flash, Recursive Flint, DeepSeek DeepInfer, Ramp Laminar, Mercury 3, and More
K3.1 Pro是国产大模型的重要能力跃迁,编程与3D能力对标国际顶尖水平,且提供完整权重与详细基准数据,值得开发者关注其技术路线与性能表现。
In today’s issue:
本期内容:
- Moonshot launches Kimi K3.1 Pro
- Anthropic previews Claude Duet
- OpenAI coordinates critical PHP fixes
- MiniMax opens M3.1 Flash weights
- Recursive ships local coding model Flint
- DeepSeek open-sources DeepInfer
- Ramp details its Laminar harness
- Inception launches diffusion LLM Mercury 3
- Databricks launches MemCurrent agent memory
- Naive RAG beats agentic memory
- Moonshot 发布 Kimi K3.1 Pro
- Anthropic 预览 Claude Duet
- OpenAI 协调关键 PHP 修复
- MiniMax 开源 M3.1 Flash 权重
- Recursive 推出本地编码模型 Flint
- DeepSeek 开源 DeepInfer
- Ramp 详解其 Laminar 框架
- Inception 发布扩散式大语言模型 Mercury 3
- Databricks 推出 MemCurrent 智能体记忆
- 朴素 RAG 胜过智能体记忆
And all the top AI dev news, papers, and tools.
以及所有顶级 AI 开发新闻、论文和工具。
Top Stories
头条故事
Kimi K3.1 Pro
Moonshot AI introduced Kimi K3.1 Pro, its most capable 3T-class model and the first in the K3.1 series, with full weights due by November 4. It is a multimodal Mixture-of-Experts model with 2.8T total and 104B active parameters and a 1M-token context, post-trained from Kimi K3 with substantially more RL compute.
Moonshot AI 推出了 Kimi K3.1 Pro,这是其最强大的 3T 级模型,也是 K3.1 系列的首款产品,完整权重将于 11 月 4 日开放。它是一个多模态混合专家(MoE)模型,总参数量为 2.8T,激活参数为 104B,上下文窗口为 1M token,并在 Kimi K3 的基础上通过大幅增加的强化学习计算量进行了后训练。
- Coding results: On Terminal-Bench 4.0, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%), and Moonshot’s charts put it first on SWE Marathon (51.7) and Agents’ Last Exam (43.2). It also builds full projects in game engines such as Unreal, Unity, and Godot.
- 3D and CAD work: K3.1 Pro writes and runs scripts for Blender and CAD software to build and edit assets. On BenchCAD, which reconstructs 3D objects from multi-view renderings through CAD code, it scores 92.7% with tools, against 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra.
- Finance and factual accuracy: It sets a state-of-the-art 64.4 on OfficeQA Pro v2 and improves on K3 by 7.2 points in financial modeling and 11.6 points in financial research on internal benchmarks. On Moonshot’s Kimi Hallu Bench Hard stress test, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%.
- Cyber controls: For the open-weight release, Moonshot removed cyber-related data from mid-training and the RL environment mix. A checkpoint without these controls solves 90.5% of its xdev-bench-hard exploit tasks at pass@3, while the final model reaches 81.0%.
- Availability: K3.1 Pro and K3.1 Pro Highspeed, which runs 2 to 3x faster, are live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
- 编码结果:在 Terminal-Bench 4.0 上,K3.1 Pro 得分 65.7%,领先于 GPT-6 Astra(58.2%),略低于 Claude Opus 5.5(66.4%)。Moonshot 的图表显示其在 SWE Marathon(51.7)和 Agents’ Last Exam(43.2)中排名第一。它还能在 Unreal、Unity 和 Godot 等游戏引擎中构建完整项目。
- 3D 和 CAD 工作:K3.1 Pro 编写并运行 Blender 和 CAD 软件的脚本以构建和编辑资产。在 BenchCAD(一种通过 CAD 代码从多视图渲染重建 3D 对象的基准测试)上,使用工具时得分为 92.7%,而 Claude Opus 5.5 为 96.2%,GPT-6 Astra 为 95.9%。
- 金融与事实准确性:它在 OfficeQA Pro v2 上创下 64.4 的最高纪录,在内部基准测试中,金融建模能力比 K3 提升 7.2 分,金融研究能力提升 11.6 分。在 Moonshot 的 Kimi Hallu Bench Hard 压力测试中,声明级别的幻觉率从 K3 的 23.4% 降至 8.3%。
- 安全控制:对于开源版本,Moonshot 在中期训练和强化学习环境混合中移除了与网络安全相关的数据。一个未包含这些控制的检查点在 pass@3 下解决了 xdev-bench-hard 漏洞利用任务的 90.5%,而最终模型达到了 81.0%。
- 可用性:K3.1 Pro 和速度提升 2 到 3 倍的 K3.1 Pro Highspeed 已在 Kimi.ai、Kimi Work、Kimi Code 和 Kimi API 上线。
Blog | API
博客 | API
Read more
阅读全文
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力