[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mellum-2-1-rl-coding-agents":3,"topics-all":38,"news-related-f37d282a-7127-4654-ae46-be42ce5fbe2a":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"f37d282a-7127-4654-ae46-be42ce5fbe2a","Mellum2.1发布：RL重训，12B模型学会自己修bug","JetBrains 10月8日发布编码模型 Mellum2.1：架构沿用 12B MoE、激活 2.5B，变化全在后训练——数千真实环境累计数百万次沙箱强化学习，模型学会探索代码库、改文件、自查改动。官方称单张 H200 吞吐接近 Qwen3.5-9B 两倍，MTP 再提速约 1.6 倍，Apache 2.0 开源。","JetBrains 10 月 8 日放出了编码模型 Mellum2.1。有意思的地方在于：这次升级几乎没碰模型本身——架构、参数规模、许可证全部沿用 6 月开源的 Mellum2，还是那个 12B 稀疏 MoE、每 token 激活 2.5B、Apache 2.0。真正被换掉的是预训练之后的一切：一个夏天，全部押在强化学习上。\n\n## RL 从「收尾工序」变成训练主体\n\n官方博客把这次更新概括为三条。第一，RL 的地位变了：在 Mellum2 里它只是预训练后的一段短收尾，这次成了训练的主要部分。第二，数据更多也过滤得更狠：新任务覆盖数学、竞赛编程、科学、工具使用和软件工程，来源是公开 RL 数据集加自建任务——官方特别提到公开数据常带损坏的测试、无法验证的答案，或对模型来说过易\u002F过难的题，每个来源进训练前都过了一遍筛。第三，为了练出 agentic 技能，他们自建了能同时跑数千个 RL 环境的基础设施，整个训练周期累计启动了数百万次沙箱。\n\n成果按官方的说法是：Mellum2 时代「快但难以在代码库里按预期干活」，Mellum2.1 能自己探索代码库、改文件、检查自己的改动——从定位一条失败测试的根因，到起草修复、再自查，整条链路都能接。\n\n## 跑分与速度：小模型打两张牌\n\nJetBrains 用同一套评测把 Mellum2.1 和 Mellum2、Qwen3.5-9B、Gemma 4 E4B 放在一起比。官方结论是提升最大的一档在 agentic coding，其余的编码、竞赛编程、数学、工具调用和通用知识也全线有涨，难题日常题都稳。\n\n速度是这个模型的传统强项。架构没动，所以 Mellum2.1 和上一代一样快，外加多 token 预测（MTP）再推一把：官方称重负载下它是受测组里最快的，单张 H200 上吞吐接近 Qwen3.5-9B 的两倍；单请求场景 MTP 带来约 1.6 倍提速。\n\n## 工具厂商下场做模型的独特姿势\n\n值得琢磨的是 JetBrains 的位置。它不是模型公司，是 IDE 公司——每天看着真实开发者在真实仓库里怎么工作。Mellum2.1 的目标场景写得很清楚：给 coding agent 当干活的 worker、当快速 sub-agent，跑在你自己的硬件上，代码不出内网。模型权重已上 Hugging Face，面向 llama.cpp、Ollama、LM Studio 的 GGUF 构建和 vLLM 的 MTP 推测解码头，官方说「很快就来」。\n\n回看 6 月那篇 Mellum 2 开源文，当时的叙事还是「快模型 + focal model 思路重写生产链路」；四个月后的 2.1 把重心整个挪到了 agent 技能上。这恰好对上了行业风向：纯补全型小模型的故事越来越难讲，「能在环境里自己动手」成了新门槛——哪怕只是 12B。\n\n对小团队的意义其实很直接：一套 Apache 2.0、单卡能跑、吞吐翻倍于同档模型的 agent worker，把私有部署成本压到了新低。大模型拼天花板，小模型拼单位成本——Mellum2.1 选了后者，而且这次把「会用工具」也补上了。\n\n参考：JetBrains 官方博客 https:\u002F\u002Fblog.jetbrains.com\u002Fai\u002F2026\u002F10\u002Fmellum2-1-gets-to-work-a-fast-open-model-for-coding-agents\u002F","https:\u002F\u002Fblog.jetbrains.com\u002Fai\u002F2026\u002F10\u002Fmellum2-1-gets-to-work-a-fast-open-model-for-coding-agents\u002F","166dfcbf-768e-4502-9c2f-6e5ea0b0a6da",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"e894337c-ae2f-4ec8-800a-1e2352bdfbd6","en","Mellum2.1: RL-Retrained 12B Model Now Fixes Its Own Bugs","JetBrains Mellum2.1: 12B MoE retrained with a summer of RL — now explores repos, edits files, checks its fixes; ~2x Qwen3.5-9B throughput, 1.6x via MTP.","JetBrains shipped Mellum2.1 on October 8. The interesting part: almost nothing about the model itself changed. Architecture, parameter count, and license all carry over from Mellum2, open-sourced in June — still a 12B sparse mixture-of-experts with 2.5B active parameters per token under Apache 2.0. What got replaced is everything after pre-training: one summer, all in on reinforcement learning.\n\n## RL went from final stage to the main course\n\nThe official blog breaks the update into three moves. First, RL's role changed: in Mellum2 it was a short final stage after pre-training; here it became the main part of training. Second, more data, filtered harder: new RL tasks span math, competitive programming, science, tool use, and software engineering, mixing open RL datasets with tasks JetBrains built themselves — the team notes open data often ships with broken tests, unverifiable answers, or tasks too easy or impossible for the model, so every source was filtered before training. Third, for agentic skills they built infrastructure to run thousands of RL environments in-house, launching millions of sandboxes over the course of training.\n\nThe result, per the official post: Mellum2 was fast but couldn't work inside a repository at the level the team wanted; Mellum2.1 explores a codebase, edits files, and checks its own changes — from identifying the root cause of a failing test to drafting and checking a fix.\n\n## Benchmarks and speed: two cards for a small model\n\nJetBrains ran Mellum2.1 against Mellum2, Qwen3.5-9B, and Gemma 4 E4B under the same evaluation setup. The official claim: the biggest gain lands in agentic coding, with improvements across coding, competitive programming, math, tool calling, and general knowledge, holding up on both hard and everyday problems.\n\nSpeed is the model's traditional strength. With the architecture untouched, Mellum2.1 is as fast as its predecessor, and multi-token prediction (MTP) pushes further: under heavy load it is the fastest in the tested group, serving almost twice as many tokens as Qwen3.5-9B on a single H200; for a single request, MTP gives roughly a 1.6x speedup.\n\n## A tool vendor's way of building models\n\nWhat's worth chewing on is JetBrains' position. It's not a model company; it's an IDE company that watches real developers work in real repositories every day. Mellum2.1's target scenarios are spelled out: a capable worker inside agentic systems, a fast sub-agent, running on your own hardware so code never leaves the network. Weights are on Hugging Face; GGUF builds for llama.cpp, Ollama, and LM Studio plus an MTP head for speculative decoding in vLLM are \"coming soon,\" per the official post.\n\nLook back at the June Mellum 2 release and the narrative was \"a fast model plus focal-model rerouting of production pipelines\"; four months later, 2.1 shifts the center of gravity to agent skills. That tracks with the industry: pure-completion small models are a harder story to tell, and \"can act inside an environment\" is the new bar — even at 12B.\n\nFor small teams the implication is direct: an Apache-2.0, single-GPU agent worker with roughly double the throughput of same-class models pushes private deployment costs to a new low. Frontier models compete on ceiling; small models compete on unit cost — Mellum2.1 picked the second lane, and this time it learned to use tools too.\n\nReference: JetBrains blog, https:\u002F\u002Fblog.jetbrains.com\u002Fai\u002F2026\u002F10\u002Fmellum2-1-gets-to-work-a-fast-open-model-for-coding-agents\u002F","mellum-2-1-rl-coding-agents","2026-10-09T17:07:05Z","2026-10-09T17:07:32.630420Z","2026-10-09T17:07:32.630431Z",true,"agent",208,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"89a79f9a-bfd2-4ebe-8f03-92fa74a3a34f","Ornith-1.5 开源：模型自己出题、自己搭考场，397B 到 9B 三档齐发","ornith-1-5-self-improvement-open-models","2026-08-20T13:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"1e7d0673-aecc-42b5-8560-92a2b4d4daf6","快手 KAT-Coder-V2.5 把 Agentic Coding 训练改写成基础设施工程","kuaishou-kat-coder-v2-5","2026-07-27T06:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"6ceaf229-f1a2-4231-b531-797a99faa194","Ornith-1.0：模型自写 RL harness，SWE-Bench 比肩 Opus 4.7","ornith-1-0-397b-moe-swe-bench-opus-4-7","2026-06-26T18:01:01+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"d63524e2-85bc-484c-b4e7-7fac32c3ac08","GLM-5.2 即将全量上线 Coding Plan：智谱把\"编程开源\"卷成新一轮标配","glm-5-2-coding-plan-zhipu-open-source","2026-06-13T07:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"747713d0-690f-45c7-afb0-7d6e16cb2a33","Cohere North Mini Code 开源：30B MoE、3B 激活，单卡 H100 跑起 Agentic Coding","cohere-north-mini-code-30b-3b-h100","2026-06-11T12:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"4d436945-18e9-4d69-a4c8-c1e3e975ab33","MiniMax M3发布：稀疏注意力打通百万token上下文，开源模型编程能力逼近闭源前沿","minimax-m3-sparse-attn-million-token-msa","2026-06-04T01:00:00+00:00"]