GPTProto
⌘K

Kimi / Moonshot AI

34 picksNewest firstLatest article Sep 24, 2026, 8:00 AM
4 stories
15:22 MarkTechPost（RSS）AI score 78/100

Fireworks AI Releases Ember-1: A Post-Trained Kimi K3 That Uses About 40% Fewer Tokens

Fireworks AI has released Ember-1, a specialized model from Fireworks Research built by post-training Moonshot AI’s open-weight Kimi K3. Ember-1 learns to produce shorter reasoning traces while keeping task accuracy. This is different from lowering the reasoning effort setting at inference time. According to the Fireworks release post, Ember-1 delivers Kimi K3’s quality with about 40% fewer tokens. Is it deployable? Yes, but only through the Fireworks serverless API as a Research Preview. Firewo
06:43 X：Elvis Saravia (@omarsar0, DAIR.AI)AI score 55/100

This is a bigger deal than it seems. I like this push on the Pareto frontier to squeeze as much as ...

This is a bigger deal than it seems.I like this push on the Pareto frontier to squeeze as much as you can out of your tokens.Feels underexplored.All labs are quick to launch models, so certain aspects just aren't optimized.These are just a few of the great things you can start doing with frontier open models.Ember-1 produces shorter reasoning traces (40% fewer tokens) without sacrificing performance.
03:09 Hacker News 热门（buzzing.cc 中文翻译）AI score 68/100

Ember-1

Ember-1: half the tokens, same answersEmber-1 is a new specialized model from Fireworks Research that delivers Kimi K3’s quality with 40% fewer tokens. Built on Kimi K3, it learned to cut unnecessary reasoning while keeping the thinking that matters. We tested it on external benchmarks, in live customer A/B tests, and on our own coding and agent workloads, and quality held up in every setting. Available today, Ember-1 kicks off an ongoing series of specialized models by Fireworks, shaped by what
01:31 Hacker News：AI 热帖AI score 64/100

Ember-1

Ember-1: half the tokens, same answersEmber-1 is a new specialized model from Fireworks Research that delivers Kimi K3’s quality with 40% fewer tokens. Built on Kimi K3, it learned to cut unnecessary reasoning while keeping the thinking that matters. We tested it on external benchmarks, in live customer A/B tests, and on our own coding and agent workloads, and quality held up in every setting. Available today, Ember-1 kicks off an ongoing series of specialized models by Fireworks, shaped by what
1 story
2 stories
1 story
19:00 X：Kimi.ai (@Kimi_Moonshot)AI score 69/100

Meet the new Kimi Browser Extension, formerly Kimi WebBridge. From your browser sidebar, you can ch...

Kimi 发布新的 Kimi Browser Extension，前身为 Kimi WebBridge。用户可在浏览器侧边栏与 Kimi 对话，让它导航网页、填写表单并完成任务；对重复性任务，可录制一次操作步骤保存为 skill，下次由 Kimi 接手执行。产品现已在 kimi.com/products/kimi-browser-extension 和 Chrome Web Store 上线。
1 story
1 story
1 story
1 story
1 story
08:00 HuggingFace Daily Papers（社区热门论文）AI score 75/100

Would You Walk to the Car Wash? Revealing the Salience Bias of Large Language Models in Commonsense Reasoning

研究提出“显著性偏差”概念，指大语言模型被输入中无用的显性干扰（如数字）劫持，忽略任务隐含的常识前提。基于新构建的 SaliTrap 基准评估 12 个主流模型，最佳模型仅 54.8% 查询避开陷阱，8/12 模型低于 30%；GLM-5.1 和 Kimi-K2 在识别陷阱后仍分别有 86.2% 和 81.8% 的遵从率。
1 story
1 story
4 stories
2 stories
1 story
1 story
1 story
1 story
1 story
1 story
00:00 Moonshot AI：Kimi BlogAI score 70/100

PerceptionBench

Moonshot AI 发布 PerceptionBench，一个从 40 多个现有基准中模型实际失败案例归纳出的视觉感知基准，包含 10 项原子感知能力和 3000 道验证题。所有测试模型准确率均未超过 60%，且大量正确答案在重复提问时无法复现，表明模型更多是猜测而非真正感知。PerceptionBench 旨在精确诊断多模态 AI 的视觉感知断裂点，推动其实现忠实、一致的视觉理解。
1 story
1 story
1 story
21:48 X：硅基流动 SiliconFlow (@SiliconFlowAI)AI score 65/100

Better Coding with Less Overthinking K2.7 Code takes K2.6's strong base and goes deep Meet @Moonshot...

硅基流动上线月之暗面Kimi K2.7 Code模型。基于K2.6改进，专注编码与智能体任务。32B激活/1T总参，VLM多模态，支持交错思考与多步工具调用。相比K2.6，推理token使用减少30%，减少过度思考；长程编码任务指令遵循和完成率提升。性能接近GPT5.5与Opus 4.8。价格：缓存输入0.19/输入0.94/输出4.00每百万token。
1 story
18:16 X：Kimi.ai (@Kimi_Moonshot)AI score 70/100

🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! 🔷 Improved coding & ...

Kimi 发布并开源最新代码模型 Kimi-K2.7-Code。相比 K2.6，其在 Kimi Code Bench v2 上提升 +21.8%，Program Bench 提升 +11.0%，MLS Bench Lite 提升 +31.5%。推理效率改进，推理 token 使用量降低 30%，长时编码任务中指令遵循和端到端成功率均提升。6x 高速模式即将推出，即日起可通过 Kimi API 和 Kimi Code 使用。
1 story
22:23 X：Kimi.ai (@Kimi_Moonshot)AI score 76/100

Meet Kimi Work - a local AI agent on your desktop that does the work for you. 🔹Native agent swarm:...

Kimi发布Kimi Work，一款本地运行的桌面AI智能体。支持最多300个AI智能体同时在本地机器并行运行；配合WebBridge浏览器扩展，智能体可在浏览器中导航、搜索、点击、填写并完成任务；内置全球市场数据工具，可直接调用Yahoo Finance和世界银行数据，无需复杂API设置；记忆系统会记录用户偏好、过往决策和上下文。支持macOS（Apple Silicon）和Windows。
1 story
1 story
08:00 Cursor BlogAI score 74/100

Introducing Composer 2.5

Cursor 平台发布了智能与行为表现大幅提升的 Composer 2.5。该模型更擅长执行复杂指令和长期任务。其改进基于训练规模的扩大、更复杂的强化学习环境及新的学习方法。关键技术包括：使用文本反馈进行针对性强化学习以纠正具体错误；采用基于真实代码库、规模达前代25倍的合成数据进行训练；并引入分片Muon优化器等新架构。模型基于Moonshot的开源检查点构建。开发团队正合作训练一个计算量十倍的更大模型，并在大规模训练中发现了新型奖励作弊问题。
1 story
00:44 X：Nathan Lambert (@natolambert)AI score 76/100

Visiting the Moonshot office was amazing, right up there with the best vibes of any AI company I’ve ...

作者盛赞访问Moonshot AI（月之暗面）的体验极佳。其旗下产品Kimi正以超200亿美元的投后估值完成新一轮约20亿美元融资，由美团龙珠领投。公司在2025年前六个月融资总额超39亿美元，估值较去年11月增长约4倍，成为中国累计融资额最高的AI初创公司。其年度经常性收入在短期内从1亿美元快速增长至超2亿美元，付费订阅和API使用量均在加速。
1 story
00:00 LMSYS：Blog（Chatbot Arena 团队）AI score 74/100

Blog Updating 1T parameters in seconds — P2P weight transfer in Large Scale Distributed RL We introduced a RDMA-based, Peer to Peer weight update mechanism for RL workloads in SGLang as a supplement to traditional NCCL broadcast methods, compatible with all major open source models. By util... Jiadong Guo, Xin Ji, Letian Ruan, Teng Ma, Chenyang Zhao, Yueming Yuan, Zhichen Zeng April 29, 2026

LMSYS团队针对SGLang中的强化学习工作负载，提出了一种基于RDMA的点对点权重更新机制，作为传统NCCL广播方法的补充。该设计利用源端CPU引擎副本和Mooncake TransferEngine进行P2P RDMA传输，将拥有1T参数的Kimi-K2模型的权重传输时间从53秒大幅缩短至7.2秒，提速7倍。其代价是每个训练等级需在CPU内存中额外占用一个32G的推理引擎副本。此优化最大限度地减少了网络冗余，允许推理服务器更快恢复rollout过程，且兼容所有主流开源模型。
1 story
20:00 Cursor BlogAI score 72/100

A technical report on Composer 2

本报告介绍了代码模型Composer 2的训练过程。该模型基于开源基础模型Kimi K2.5，通过两阶段训练：首先进行侧重代码的持续预训练以深化编码知识，随后在高度模拟真实Cursor环境的大规模强化学习中提升端到端智能体性能。在自建的真实任务评估集CursorBench上，Composer 2得分为61.3，较前代提升37%，与前沿模型性能相当。在公开基准SWE-bench Multilingual和Terminal-Bench上分别获得73.7和61.7分，并在保持高精度的同时实现了显著更低的推理成本。训练依托为Blackwell GPU定制的高效MoE训练内核、跨区域异步强化学习管道等大规模基础设施完成。
1 story
1 story
00:00 Moonshot AI：Kimi Blog

Kimi Vendor Verifier

Kimi 发布 K2.5 模型时开源 Kimi Vendor Verifier（KVV），用于验证第三方推理实现的准确性。针对开源模型部署渠道多样化导致的质量失控问题，KVV 提供六项关键基准测试，覆盖参数约束验证、多模态流水线、长输出压力测试、工具调用一致性及编程能力评估。项目与 vLLM/SGLang 社区合作修复根因，并提供预发布验证和实时更新的公开排行榜。完整评估在双 H20 8 卡服务器上约需 15 小时。
1 story
1 story
1 story