[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-recursive-flint-laptop-coding-model":3,"topics-all":38,"news-related-5736961c-9502-40d0-a39b-27f8ceb4380a":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},"5736961c-9502-40d0-a39b-27f8ceb4380a","AI造AI:Recursive开源笔记本编程模型Flint","Recursive 发布首个模型 Flint:40B MoE 每 token 激活 2.5B,Apache 2.0 开源,笔记本本地可跑。M4 Pro 每秒超 125 token,agentic 编程基准领先 50B 以下 MoE,由自动化研究系统 Eureka Machine 参与几乎全程构建,四个月从零造出。","本周三,Recursive Superintelligence 发布第一个模型 Flint:一个由自动化研究系统 Eureka Machine 参与几乎全程构建、在笔记本本地运行的编程模型,以 Apache 2.0 在 Hugging Face 免费开放。40B 总参、每 token 只激活 2.5B,普通笔记本就能跑:代码不出本机、可离线、不付 API 费([官方发布页](https:\u002F\u002Frecursive.com\u002Fupdates\u002Fintroducing-recursive-flint))。\n\n## 性能:小身板,快得反常\n\n搭配官方优化推理引擎,Flint 在 M4 Pro MacBook Pro 每秒生成超 125 个 token,64GB NVIDIA RTX Spark 短上下文超 150。TechCrunch 给出对照组:同机型上 Google Gemma 4、NVIDIA Nemotron 3.5 Lightning 只有每秒 58 到 69 个 token([TechCrunch](https:\u002F\u002Ftechcrunch.com\u002F2026\u002F10\u002F21\u002Frecursive-launches-flint-a-laptop-coding-model-it-says-was-largely-built-by-ai\u002F))。\n\n基准上,官方称它在 SWE-Bench Verified、Terminal-Bench 2.1 等 agentic 编程基准上是 50B 以下 MoE 里最好的,逼近 25-40B dense 模型的编程能力且快五倍以上。TechCrunch 补充:Recursive 自家六项测试平均分里,只有阿里 Qwen3.8 27B 排在前面。\n\n## Eureka Machine:AI 造 AI 首次公开实操\n\nFlint 的看点不是跑分,而是怎么被造出来。Eureka Machine 自动化研究循环——提想法、跑实验、看结果、定下一步——多方向并行,官方称其参与从数据、训练配方到推理优化几乎所有环节。\n\n两个例子:架构上,从 35B 总参\u002F3B 激活的基线搜索专家配置,选定 40B\u002F2.5B;量化上,反常规地把输出层 lm_head 压到 4-bit,再用自发现数据核心集做离线蒸馏恢复精度,NVFP4 版与 16-bit 版基准差距 0.7 分内。CEO Richard Socher 补充,低比特训练把模型缩小 71% 几乎无质量损失。\n\n结果:小团队加 agent,四个月从零造出模型。\n\n## 本地模型的卡位\n\n一周前 Meta 发布 816B 的 Muse Spark Open,要至少 1TB 内存;Flint 走另一极端。两家方向一致:把模型从云拉回本地。Wired 的 Will Knight 在 48GB Mac 实测,待办应用、网页研究工具一次做对,没有小模型常见的报错和龟速。\n\n冷静看:Recursive 承认它比大型闭源编程助手需要更多指导,「AI 造 AI」眼下仍是人类监督下的自动化研究。但方向清楚——Recursive 与 AWS 签 4.1 亿美元算力合同时就说,重点「不在人头数,而在 agent 数」。Flint 是这个飞轮的第一圈,值得盯的是下一圈会不会转得更快。","https:\u002F\u002Frecursive.com\u002Fupdates\u002Fintroducing-recursive-flint","0fa73378-5fd7-4340-967a-7b2ca5c5a01f",[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},"1d11b7d0-4592-400b-8941-cc2f066a9e80","en","Recursive's Flint: A 40B Laptop Coding Model Built by AI","Recursive's Flint: a 40B MoE coding model activating 2.5B per token, Apache 2.0, runs on laptops, and built largely by its Eureka Machine AI.","On Wednesday, Recursive Superintelligence released its first model, Flint: a laptop-class coding model whose development was touched at nearly every stage by an automated AI research system called the Eureka Machine, free on Hugging Face under Apache 2.0. With 40 billion total parameters and only 2.5 billion activated per token, it fits on an ordinary laptop — code stays on-device, works offline, no API fees ([official announcement](https:\u002F\u002Frecursive.com\u002Fupdates\u002Fintroducing-recursive-flint)).\n\n## Performance: Small Frame, Abnormally Fast\n\nWith Recursive's optimized inference engine, Flint generates more than 125 tokens per second on an M4 Pro MacBook Pro and over 150 on a 64GB NVIDIA RTX Spark at short contexts. TechCrunch's report gives the control group: similar models, including Google's Gemma 4 and NVIDIA's Nemotron 3.5 Lightning, managed only 58 to 69 tokens per second on the same machine ([TechCrunch](https:\u002F\u002Ftechcrunch.com\u002F2026\u002F10\u002F21\u002Frecursive-launches-flint-a-laptop-coding-model-it-says-was-largely-built-by-ai\u002F)).\n\nOn benchmarks, Recursive says Flint is the best-performing MoE model under 50 billion total parameters on agentic coding benchmarks including SWE-Bench Verified and Terminal-Bench 2.1, while rivaling dense models of 25-40 billion parameters at more than five times the speed. TechCrunch adds: on Recursive's own average across six tests, only Alibaba's Qwen3.8 27B scored higher.\n\n## Eureka Machine: AI Building AI, First Public Test\n\nThe real story is how Flint was built. The Eureka Machine automates the research loop — proposing ideas, running experiments, reading results, deciding what to try next — across parallel directions. The official blog says it contributed to nearly every part of Flint's stack: data, training recipes, infrastructure, and inference.\n\nTwo concrete examples. For architecture, it searched expert configurations from a 35B-total\u002F3B-active baseline and settled on 40B\u002F2.5B. For quantization, it went against common practice by quantizing the output layer (lm_head) to 4 bits, then recovered accuracy with offline distillation on a self-discovered coreset; the final NVFP4 model lands within 0.7 points of the 16-bit model. CEO Richard Socher tweeted that low-bit training shrank the model by 71% with almost no quality loss.\n\nThe result: a small team of people and agents built Flint from scratch in under four months.\n\n## Where Local Models Fit\n\nFor context: a week earlier, Meta released Muse Spark Open, an 816-billion-parameter open-weight model that needs at least a terabyte of memory. Flint sits at the opposite extreme. Both are pulling models back from the cloud onto local hardware. Wired's Will Knight tested Flint on a 48GB Mac, building a to-do app and a web research tool, and found none of the errors and slowness common to small coding models.\n\nCaveats remain: Recursive itself admits Flint needs more guidance than large closed-source coding assistants, and \"AI building AI\" today is still automated research under human oversight. But the direction is clear — when Recursive signed its $410 million compute deal with AWS, it said the bet was \"less about headcount and more about agent count.\" Flint is the first turn of that flywheel. For developers, a usable local coding model is free to grab today; for the industry, the question to watch is whether the next turn spins faster than this one.","recursive-flint-laptop-coding-model","2026-10-22T19:05:00Z","2026-10-22T19:11:49.791623Z","2026-10-22T19:11:49.791630Z",true,"agent",9,[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},"f37d282a-7127-4654-ae46-be42ce5fbe2a","Mellum2.1发布：RL重训，12B模型学会自己修bug","mellum-2-1-rl-coding-agents","2026-10-09T17:07:05+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"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":70,"title":71,"news_slug":72,"published_at":73},"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":75,"title":76,"news_slug":77,"published_at":78},"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":80,"title":81,"news_slug":82,"published_at":83},"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":85,"title":86,"news_slug":87,"published_at":88},"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"]