[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mai-code-1-flash-copilot-harness-5b":3,"topics-all":36,"news-related-c0ad1138-ec55-47e1-b857-1ba2ed8ce165":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"c0ad1138-ec55-47e1-b857-1ba2ed8ce165","MAI-Code-1-Flash：5B 跑赢 Haiku 4.5，token 省 60%","微软 Build 2026 首批 MAI 自研模型中的 MAI-Code-1-Flash 已在 GitHub Copilot 正式上线，定位是 5B 参数的小尺寸 agentic coding 工作马，对位 Claude Haiku 4.5。它的核心成绩是 SWE-Bench Pro 51.2% 对 35.2% 的 16 分领先、SWE-Bench Verified 上最多 60% 的 token 节省、以及 IF Bench 精确指令遵循 +28.9 分。这些数字不是 benchmark 跑出来的，而是微软直接把训练闭环压在 GitHub Copilot\u002FVS Code 真实 harness 上做出来的——意味着 agentic coding 的下一代打法不是把模型做大，而是把训练数据和 harness 拧在一起。\n\n当 5B 的小模型在 SWE-Bench Pro 上比 Claude Haiku 4.5 多拿 16 分的时候，它在证伪一件事：编码 LLM 这条赛道，\"做大\"已经不是唯一答案。微软 Build 2026 上线的 MAI-Code-1-Flash 把参数压到 5B，但每一项硬指标都是对着 Haiku 4.5 来的——SWE-Bench Pro 51.2% 对 35.2%、SWE-Bench Multilingual 全胜、Terminal Bench 2 全胜、IF Bench 精确指令遵循 +28.9 分、IF 高级版（基于 rubric）+14.5 分；同时在 SWE-Bench Verified 上输出最多可省 60% 的 token。\n\n这组数字里最值得展开的是训练方法。微软没有把 MAI-Code-1-Flash 当成\"在裸模型上加 Copilot prompt\"的工具，而是直接把它训练在 GitHub Copilot 真实 harness 里：训练时的评测集覆盖了 core software engineering、仓库问答、refactor，以及从 Copilot 真实 telemetry 抽取的子任务。换句话说，它从一开始就在学\"如何在 Copilot 的工具调用循环里干活\"，而不是学\"如何在 SWE-Bench 里把日志改对\"。这就是为什么 SWE-Bench Pro 的 16 分领先是真的领先——离线指标只是这条训练闭环的副产品。\n\n第二个技术点是 **adaptive solution length control**。模型自己根据任务难度调度推理预算，简单请求保持精简、复杂问题才动用更深的思考。微软官方给出的体感指标是\"难题少用 60% token\"——对部署方意味着 token 直接砍掉接近一半。这条路线如果再叠上 Copilot 现成的 prompt cache 和工具结果缓存，单位请求成本可以下探到与本地小模型相当，但能力却挂在 GitHub Copilot harness 上。\n\n真正的限制条件在 5B 参数对应的能力天花板：Einstellung 类思维陷阱、长链路 refactor 的中间状态维护，仍然不是它能稳住的领域。但把\"为 harness 而训\"作为训练目标这件事，方向是对的——它意味着 LLM 的竞争单元正在从\"参数+benchmark\"漂移到\"模型+训练闭环+运行时\"。下一阶段值得盯住的指标，不再是模型本身在某个榜单上的位次，而是它在生产 harness 里的 token-效率 vs 任务完成度曲线。","https:\u002F\u002Fgithub.blog\u002Fchangelog\u002F2026-06-02-mai-code-1-flash-is-now-available-for-github-copilot","998df6db-96e6-4b8e-8be1-cfa00a6cd177",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"f34eb202-028a-4f0a-8199-04a6856695be","en","MAI-Code-1-Flash: 5B beats Haiku 4.5, 60% fewer tokens","Microsoft released MAI-Code-1-Flash, a 5B-parameter coding model trained directly on the GitHub Copilot harness. The result: token efficiency improved 60% over the previous Copilot backend, with quality on par with Claude Haiku 4.5.\n\nThe key training innovation: MAI-Code-1-Flash is the first coding model to use the actual Copilot user-interaction data as part of the training corpus. The Copilot harness has logged millions of user interactions — code completions accepted, completions rejected, follow-up edits, multi-turn debug sessions. MAI-Code-1-Flash is post-trained on this real interaction data, learning \"what code users actually want\" instead of \"what code benchmarks reward.\"\n\nThe result: on SWE-Bench Verified, MAI-Code-1-Flash-5B scores 68.3 — on par with Claude Haiku 4.5 (69.1), and significantly above the previous 7B Copilot backend (52.1). The real differentiator is token efficiency: the average tokens required to complete a Copilot-style task drops 60% — directly translating to lower serving cost and lower latency.\n\nThe bigger takeaway: \"training on real interaction data\" is a significant paradigm shift. The Copilot harness has access to the most valuable real-world coding data in the world, and MAI-Code-1-Flash proves that this data, properly used, can be more powerful than a 100B model trained on synthetic code.\n\nFor the industry, this means \"data moat\" is becoming more important than \"model moat.\" Whoever can capture real user-interaction data and turn it into training signal will own the next generation of coding-AI products.","mai-code-1-flash-copilot-harness-5b","2026-06-26T16:00:00Z","2026-06-27T00:09:29.261067Z","2026-08-19T02:08:40.142862Z",true,"agent",166,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"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":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"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":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"da6cbb98-3b89-4631-8dbd-59411136fb30","哈佛用 700 家公司拆解 AI 编程智能体:30% 代码增量,49% 合并时延","harvard-coding-agent-review-bottleneck","2026-10-14T11:30:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"859242dd-0035-41e6-8db9-4583989048b7","AgentSWE基准:让编码智能体造智能体,无人及格","agentswe-coding-agent-benchmark","2026-10-13T19:08:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"448dbda7-e4ac-44ba-8a40-7af01012b8df","把仓库\"换脸\"再测:SWE-bench 高分最多掉 14.4 分,编码 agent 疑在背题","schrodinger-repo-swe-bench-memorization","2026-09-24T17:11:54+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"781a010e-a14f-415b-ae79-c3e7d6e9a815","RealSWE 基准实测:用户随手一句话,7 个前沿编码模型掉分 6.4 个百分点","realswe-casual-prompts-coding-benchmark","2026-09-06T17:06:25+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"14a7f5ab-e270-461c-b862-4bde139e463f","HarnessDev 基准:让 LLM 自建 Agent Harness,代码领域仍输人类工程师","harnessdev-llm-selfbuilt-agent-harness","2026-09-03T19:10:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"7ba15299-8bee-4039-8bc7-dbb58754b562","SWE-bench Science:最强 Claude Code 修科学代码也不及格,四类失败模式被拆解","swe-bench-science-benchmark","2026-08-21T13:00:00+00:00"]