[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-poolside-laguna-s-2-1-swe-bench":3,"topics-all":37,"news-related-52dce80b-183a-440b-9d47-db79cc535c90":56},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":30,"published_at":31,"created_at":32,"modified_at":33,"is_published":34,"publish_type":35,"image_url":14,"view_count":36},"52dce80b-183a-440b-9d47-db79cc535c90","Poolside Laguna S 2.1:9 周训练的 118B MoE 把 SWE-Bench 顶到 78.5%","7 月 21 日,Poolside 把 Laguna S 2.1 全部权重推到 Hugging Face 开源(OpenMDW-1.1 协议),同步发了官方 GGUF、MLX、INT4、FP8、NVFP4 量化版和 DFlash 推测解码草稿模型。架构是 118B 总参、每 token 激活 8B 的 MoE,256 路由专家里选 10 个再加 1 个共享专家,48 层、1:3 全局 \u002F 滑动窗口混合注意力,1M 上下文。\n\n最反常识的成绩单:SWE-Bench Multilingual 78.5%,Terminal-Bench 2.1 70.2%,DeepSWE v1.1 40.4%。8B 激活跑出 78.5% multilingual SWE-bench,意味着它能在单台 NVIDIA DGX Spark(桌面级 Grace Blackwell)上拿到和 Tencent Hy3 295B-A21B 接近的水平。Hermes Agent、pi、OpenCode、OpenClaw、Cline 当天就把 S 2.1 接进默认模型列表——这是「agent 时代尺寸曲线必须重新画」的明确信号。\n\n让人议论的不是学术分数,是工程:5 月 22 日开训,7 月 21 日发布,9 周;409k agentic \u002F 非 agentic 训练环境,后训练第一次用 FP8 RL 加速;rollout 预算拉到「单任务几万到几十万 token、几百步」级别;同一 prompt 在多个 harness 上并行 rollout,防止只适配自家脚手架。Poolside 公开了所有最终评测轨迹,最出圈的一段:模型在无人监督的沙箱里自己重发现了 Erdős #397 猜想——一道 2026 年 1 月才被 GPT-5.2 Pro 解决的 50 年悬案,而且给出的 8 指数线性族与公开的 6 指数解结构不同。\n\n更耐看的是它改自己 harness 的 case:Poolside 把自家 agent 框架喂回去,20 轮循环迭代,最后整体提速 5.2%、内存分配下降约 70%。他们在 blog 里写得很直白:「我们没加更多 intelligence,我们加的是『验证、不放弃、不宣布胜利』的工程行为。」\n\n服务侧:OpenRouter 256K 免费 + 1M 付费端点(定价 $0.10 入 \u002F $0.20 出 \u002F $0.01 cache 读 \u002F M token),Baseten、Vercel AI Gateway 同步上线。vLLM、SGLang、Ollama、TRT-LLM day-0 支持。\n\n我的判断:S 2.1 是 2026 下半年「小而精的 agentic coding 模型」路线的集中体现。MoE + 8B 激活 + 强 RL + 多 harness 训练 + DFlash 推测解码——这套组合拳比单纯堆参数更值得复刻。Poolside 下一档 Laguna 已经开预训练,但 S 2.1 的意义在于:它把「单机跑前 10% agent 编码能力」从 PPT 变成了可下载的权重。\n\n当 8B 激活就能在 SWE-Bench 上摸到 70%+,我们还需要 1T+ 总参的闭源大模型来写代码吗?","7 月 21 日,Poolside 把 Laguna S 2.1 全部权重推到 Hugging Face 开源(OpenMDW-1.1 协议),同步发了官方 GGUF、MLX、INT4、FP8、NVFP4 量化版和 DFlash 推测解码草稿模型。架构是 118B 总参、每 token 激活 8B 的 MoE,256 路由专家里选 10 个再加 1 个共享专家,48 层、1:3 全局 \u002F 滑动窗口混合注意力,1M 上下文。\n\n最反常识的成绩单:SWE-Bench Multilingual 78.5%,Terminal-Bench 2.1 70.2%,DeepSWE v1.1 40.4%。8B 激活跑出 78.5% multilingual SWE-bench,意味着它能在单台 NVIDIA DGX Spark(桌面级 Grace Blackwell)上拿到和 Tencent Hy3 295B-A21B 接近的水平。Hermes Agent、pi、OpenCode、OpenClaw、Cline 当天就把 S 2.1 接进默认模型列表——这是「agent 时代尺寸曲线必须重新画」的明确信号。\n\n让人议论的不是学术分数,是工程:5 月 22 日开训,7 月 21 日发布,9 周;409k agentic \u002F 非 agentic 训练环境,后训练第一次用 FP8 RL 加速;rollout 预算拉到「单任务几万到几十万 token、几百步」级别;同一 prompt 在多个 harness 上并行 rollout,防止只适配自家脚手架。Poolside 公开了所有最终评测轨迹,最出圈的一段:模型在无人监督的沙箱里自己重发现了 Erdős \n更耐看的是它改自己 harness 的 case:Poolside 把自家 agent 框架喂回去,20 轮循环迭代,最后整体提速 5.2%、内存分配下降约 70%。他们在 blog 里写得很直白:「我们没加更多 intelligence,我们加的是『验证、不放弃、不宣布胜利』的工程行为。」\n\n服务侧:OpenRouter 256K 免费 + 1M 付费端点(定价 $0.10 入 \u002F $0.20 出 \u002F $0.01 cache 读 \u002F M token),Baseten、Vercel AI Gateway 同步上线。vLLM、SGLang、Ollama、TRT-LLM day-0 支持。\n\n我的判断:S 2.1 是 2026 下半年「小而精的 agentic coding 模型」路线的集中体现。MoE + 8B 激活 + 强 RL + 多 harness 训练 + DFlash 推测解码——这套组合拳比单纯堆参数更值得复刻。Poolside 下一档 Laguna 已经开预训练,但 S 2.1 的意义在于:它把「单机跑前 10% agent 编码能力」从 PPT 变成了可下载的权重。\n\n当 8B 激活就能在 SWE-Bench 上摸到 70%+,我们还需要 1T+ 总参的闭源大模型来写代码吗?","https:\u002F\u002Fpoolside.ai\u002Fblog\u002Fintroducing-laguna-s-2-1","58615d2f-a323-4efd-8721-543ea7cf4e2c",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",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":29},"d8edfcef-5fba-45db-a393-1bfa2f433678","en","Poolside Laguna S 2.1: nine weeks, 118B MoE, 78.5% SWE-Bench","On July 21, Poolside open-sourced the full weights of Laguna S 2.1 to Hugging Face (under the OpenMDW-1.1 license), along with official GGUF \u002F MLX \u002F INT4 \u002F FP8 \u002F NVFP4 quantized variants and a DFlash speculative-decoding draft model. The architecture is a 118B-total \u002F 8B-active MoE, picking 10 of 256 routed experts plus 1 shared expert per token, with 48 layers, a 1:3 global \u002F sliding-window mixed attention layout, and 1M context. The most counter-intuitive scoreboard: SWE-Bench Multilingual 78.5%, Terminal-Bench 2.1 70.2%, DeepSWE v1.1 40.4%. Hitting 78.5% on multilingual SWE-bench with only 8B active means it can run on a single NVIDIA DGX Spark (desktop-class Grace Blackwell) and still reach a level close to Tencent Hy3 295B-A21B. On launch day, Hermes Agent, pi, OpenCode, OpenClaw and Cline all put S 2.1 into their default model lists — an unmistakable signal that the size curve for the agent era has to be redrawn. What's getting attention isn't the academic score, it's the engineering: training started May 22 and shipped July 21 — 9 weeks; 409k agentic and non-agentic training environments; FP8-accelerated RL for the first time in post-training; rollout budgets pushed to \"tens of thousands of tokens and hundreds of steps\" per task; the same prompt rolled out in parallel across multiple harnesses to prevent overfitting to any one scaffold. Poolside published every final evaluation trace, and the most circulated one is the model independently rediscovering Erdős conjecture #397 in an unsupervised sandbox — a 50-year-old open problem first solved in January 2026 by GPT-5.2 Pro, and the model's 8-exponent linear family is structurally different from the public 6-exponent solution. Even more interesting is the case of the model editing its own harness: Poolside fed its agent framework back to itself, ran 20 rounds of iterative refinement, and ended with a 5.2% overall speedup and roughly 70% reduction in memory allocation. The blog puts it bluntly: \"We didn't add more intelligence; we added the engineering behavior of 'verify, don't give up, don't declare victory'.\" On the serving side, OpenRouter offers 256K free + 1M paid endpoints (priced at $0.10 in \u002F $0.20 out \u002F $0.01 cache-read per million tokens); Baseten and Vercel AI Gateway went live the same day. vLLM, SGLang, Ollama and TRT-LLM all support it day-0. My take: S 2.1 is the concentrated expression of the \"small but precise\" agentic-coding path in the second half of 2026. MoE + 8B active + strong RL + multi-harness training + DFlash speculative decoding — this combination is more worth replicating than simply stacking parameters. Poolside's next-tier Laguna is already in pretraining, but S 2.1's meaning is that it turns \"single-machine top-10% agentic coding\" from a slide into downloadable weights.","poolside-laguna-s-2-1-swe-bench","2026-07-22T12:00:00Z","2026-07-22T14:06:51.360816Z","2026-08-19T02:08:40.142862Z",true,"agent",252,[38,47],{"slug":39,"tag_slug":39,"title_zh":40,"title_en":41,"intro_zh":42,"intro_en":43,"id":44,"is_active":34,"created_at":45,"modified_at":46},"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":48,"tag_slug":48,"title_zh":49,"title_en":50,"intro_zh":51,"intro_en":52,"id":53,"is_active":34,"created_at":54,"modified_at":55},"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":57},[58,63,68,73,78,83],{"id":59,"title":60,"news_slug":61,"published_at":62},"48726cf0-6f0c-4b09-841e-06aa8b0ffcfd","上海AI Lab Intern-Decision:多模态决策模型,4B 微超 Jev 90.02%","intern-decision-qwen3-5-multimodal-decision-model","2026-10-18T11:13:06+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"4cc278cf-6b53-4722-81ac-d1909b50e6a9","Meta 开源 816B 模型，Mac 可本地跑","meta-muse-spark-open-816b-mac","2026-10-15T23:10:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"7a2aa1aa-74ec-494a-a828-e6c1ad5b4cfe","Cloudflare Clef 决策模型开源,Jev 被压制","cloudflare-clef-open-source-decision-model-jev","2026-10-02T09:00:00+00:00",{"id":74,"title":75,"news_slug":76,"published_at":77},"71dde565-b87a-4470-83a5-0ad7a4ee1787","IBM Granite 4.2 开源:原生思维链做成开关,30B 拿下 SWE Bench Pro","ibm-granite-4-2-native-reasoning-agents","2026-08-28T14:00:00+00:00",{"id":79,"title":80,"news_slug":81,"published_at":82},"f26ace13-9c96-47ea-a528-b6682a22aa1e","Apodex 1.1 把推理搬进真实执行:PIVOT-RL 定位关键决策点,35B mini 开源","apodex-1-1-agentic-execution-pivot-rl","2026-08-25T14:30:00+00:00",{"id":84,"title":85,"news_slug":86,"published_at":87},"b4754043-6b19-499f-8459-f8fc786f4d80","Pokee-Isaac 28B 把 10M 上下文塞进客户边界:28B 参数在 RULER 10M 上 93.3%","pokee-isaac-28b-10m-context","2026-08-20T14:00:00+00:00"]