[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-reflection-beam-501b-open-weights":3,"topics-all":38,"news-related-f8725ca7-b0b9-430f-9c63-dabe9a2c1d57":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},"f8725ca7-b0b9-430f-9c63-dabe9a2c1d57","Beam 501B官宣:23B激活参数打效率牌","Reflection AI 结束隐身,首发 501B 参数稀疏 MoE 模型 Beam,23B 激活主打推理效率:官方称先进推理基准对标 GLM 5.2 时计算量省 3-4 倍,1.05 万张 GB300 跑四周生成超 1 亿条 rollout。权重本月底以 Apache 2.0 放出,目前仍在最终红队评测。","一家融资超 40 亿美元、估值 250 亿的 AI 公司,长期隐身、此前没有任何公开产品——10 月 5 日,它交出的第一份答卷是一个 501B 参数的开源模型。Reflection AI 在[官方博客](https:\u002F\u002Freflection.ai\u002Fblog\u002Fintroducing-beam)官宣首个开放权重模型 Beam:501B 总参数的稀疏 MoE,每 token 只激活 23B,定位编码、推理与智能体负载。HN 评论区有人按公开资料点破:这是这家隐身公司的第一个公开产品。\n\n## 官宣里最硬的数字\n\n预训练用了 23.8 万亿 token,来自网页与专有授权数据集。更有信息量的是强化学习:官方用 1.05 万张 NVIDIA GB300 连训四周,生成超 1 亿条 rollout,最大上下文 256K,消耗约 13 亿个沙箱。官方自称\"开放实验室迄今规模最大的 RL 训练之一\"——自报口径。参照:Inkling 3000 万条 rollout、MiMo 75.3 万条。\n\n工程细节值得单讲:Beam 用全异步策略梯度训练,官方称样本比当前策略落后一天、隔着 107 个权重版本,数值依然稳定。大规模异步 RL 的陈旧性问题被工程化解决,这可能比跑分更值钱。\n\n## 跑分:不是全能王,也没装\n\n官方跑分表覆盖编码、智能体、推理、通识四类,关键位次:DeepSWE v1.1 上 Beam 44.4、GLM 5.2 是 44.0、Kimi K3 68.0;Terminal Bench v2.1 上 80.1,略低于 GLM 5.2 的 81.0;SWE Bench Pro v2-Hard 77.2,高于 Inkling 的 56.9,低于 GLM 5.3 的 84.3;AIME 2026 97.8 对 GLM 5.2 的 99.2;HLE(无工具)36.2,落后 Kimi K3 的 46.9。\n\n官方定性也直白:编码与智能体上对标 GLM 5.2、逼近 Qwen 3.8-Max,\"Kimi K3 在原始能力上仍然领先\"——这份发布稿没冲跑分第一。\n\n## 卖点:每个 token 的推理账\n\nBeam 的主张是推理效率——官方称在先进推理基准上取得与 GLM-5.2 相当的成绩时,推理计算量省 3-4 倍;对比 Qwen 3.8-Max 这类 2T+ 家族差距更大。原理不神秘:23B 激活天然比 2T 级便宜,官方按\"2 × 激活参数 × 生成 token 数\"估 FLOPs,对照数据来自 Artificial Analysis 与 DataCurve。\n\n模型带 reasoning effort 参数:低档短回答省 token,高档长推理冲性能。RL 阶段用可控长度惩罚,先学会少说话把题做对,再学会该长则长。另一个泛化观察:RL 混合里没有浏览任务,浏览能力却随编码任务一起涨;给了 web 权限,模型自己学会调用其他大模型和 OCR API。\n\n## 权重月底才放,现在别急着排队跑分\n\n注意发布状态:Beam 还在最终红队中,早期版本只对 waitlist 部分用户开放;权重、技术报告、模型卡与工具栈本月底以 Apache 2.0 放出。聚合站 theopenweights 已标它为\"可下载权重\",与官方口径不符,以官方博客为准。\n\n安全侧用 deliberative alignment 把安全政策写进模型推理,官方称奖励预测相关性 r=0.79,并承诺开源安全评测套件。\n\n## 所以呢\n\n开源竞赛的记分牌正在换:从\"谁的参数更大\"转向\"谁的同能力推理更便宜\"。501B 总参、23B 激活,把\"旗舰能力、中型成本\"走成明牌——对自托管企业,每 token 成本比榜单第一重要得多。权重真落地、第三方复现跑分那天,才是给这家隐身公司打分的日子。","https:\u002F\u002Freflection.ai\u002Fblog\u002Fintroducing-beam","b0384237-a532-4411-9665-f40d81178241",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"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},"f96858d0-739e-4a13-af1a-2670f8ef8816","en","Reflection's Beam: 501B MoE Bets on Inference Efficiency","Stealth Reflection AI ships Beam: 501B MoE, 23B active, 100M+ rollouts. Coding near GLM 5.2 at 3-4x less compute; weights land this month.","A company that raised over $4 billion at a $25 billion valuation while shipping nothing publicly has finally shown its hand. On October 5, Reflection AI announced [Beam](https:\u002F\u002Freflection.ai\u002Fblog\u002Fintroducing-beam), its first open-weight model: a sparse Mixture-of-Experts system with 501 billion total parameters and 23 billion active per token, built for coding, reasoning, and agentic workloads. A Hacker News commenter flagged just how unusual that is: per public sources, this stealth outfit had never released a single public product until now.\n\n## The hard numbers\n\nBeam was pretrained on 23.8 trillion tokens from web and proprietary licensed datasets. The more interesting part is reinforcement learning: Reflection ran a four-week high-compute RL campaign on 10.5K NVIDIA GB300 GPUs, generating more than 100 million rollouts with a maximum context length of 256K. Training and grading consumed roughly 1.3 billion sandboxes, backed by a pool of one million coding, agentic, and STEM environments. The company calls it \"one of the largest scale RL runs conducted by any open lab to date\" — their own framing, worth noting. For scale, Reflection's comparison puts Inkling at 30M rollouts and MiMo at 753K.\n\nOne engineering detail stands apart: Beam was trained with fully asynchronous policy gradients. Reflection claims numerics stayed stable even when samples were generated more than a day earlier — 107 weight versions behind the current policy. Solving staleness at that scale may matter more than any single benchmark score.\n\n## Benchmarks: honest, not triumphant\n\nThe published tables span coding, agentic, reasoning, and general benchmarks:\n\n- DeepSWE v1.1: Beam 44.4 vs GLM 5.2 at 44.0, Kimi K3 at 68.0, DeepSeek V4.1 Flash at 74.2\n- Terminal Bench v2.1: Beam 80.1, just under GLM 5.2's 81.0; Kimi K3 88.3, DeepSeek V4.1 Flash 90.6\n- SWE Bench Pro v2-Hard: Beam 77.2, well above Inkling's 56.9, below GLM 5.3's 84.3 and Kimi K3's 88.2\n- AIME 2026: Beam 97.8 vs GLM 5.2's 99.2\n- HLE without tools: 36.2, behind GLM 5.3's 42.3 and Kimi K3's 46.9\n\nReflection's own framing is blunt: competitive with GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks, while \"Kimi K3 remains ahead on raw capability.\" No pretend crown here.\n\n## The actual pitch: inference efficiency\n\nBeam's real claim is efficiency — scores comparable to GLM-5.2 on advanced reasoning benchmarks at 3–4× less inference compute, with even bigger gaps against 2T+ families like Qwen 3.8-Max. The mechanism is unglamorous: 23B active parameters are simply cheaper per token than 2T-class models. Reflection estimates FLOPs as roughly 2 × active parameters × generated tokens, drawing comparison data from Artificial Analysis and DataCurve.\n\nThe model ships with a reasoning-effort parameter: lower settings bias toward shorter answers, higher settings allow longer reasoning. During RL, a controllable length penalty taught it to solve tasks with fewer tokens first, then expand reasoning where it paid off. One notable generalization: browsing ability improved even though browsing tasks were absent from the RL mixture; given web access, the model organically learned to search and query other LLMs and use OCR APIs to read documents.\n\n## Weights land later this month\n\nMind the release state: Beam is still in final red-teaming, with early access limited to a waitlist. Weights, technical report, model card, and a full stack for running, evaluating, and fine-tuning arrive \"later this month\" under Apache 2.0 — the official blog's wording, which trumps aggregator pages already tagging Beam as \"downloadable.\"\n\nOn safety, Reflection used deliberative alignment — baking the safety policy into the model's reasoning — and says its reward forecasting hit r = 0.79 versus 0.46 for a Best-of-N ceiling. The company promises to open-source the safety evaluation suite it built and used internally.\n\n## So what\n\nThe open-weights scoreboard is shifting from \"biggest parameter count\" to \"cheapest inference at frontier capability.\" A 501B-total \u002F 23B-active MoE is the most explicit bet yet that enterprises care more about cost per token than leaderboard rank — a workhorse pitch, not a trophy pitch. Circle the end of this month: the day independent teams can actually download the weights and re-run these numbers is the day this stealth company gets a real grade.","reflection-beam-501b-open-weights","2026-10-06T13:13:51Z","2026-10-06T13:13:54.637071Z","2026-10-06T13:13:54.637083Z",true,"agent",364,[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},"d941056b-c2e7-42e5-965a-a982c20b1169","Qwen3.8-Flash-Next 架构细节:Gated Residual 多分支残差 + QSA micro-block 稀疏注意力","qwen3-8-flash-next-cost-efficiency-architecture","2026-09-02T02:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"08872b65-a25f-41f2-877c-a6f265e3c72f","LFM2.5-24B开源:2B激活跑赢Qwen3.6-35B","liquid-ai-lfm2-5-24b-a2b","2026-10-14T23:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"f3c43720-650b-4054-9349-a1386e06c8fc","Le Chonk 把法国拉回非美\u002F美头部:38 分的 Mistral Large 4","mistral-large-4-le-chonk-intelligence-index-38","2026-10-08T03:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"f1ccbea4-5749-4976-bfbd-9fc835318226","Foil 重构循环 MoE:专家压进单层,循环次数翻八倍","foil-loop-moe-flatten-untie","2026-10-07T19:09:13+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"5ddba2a9-781e-4763-b1e0-1e20c6480391","Mistral Large 4:1万亿参数MoE,月底开源","mistral-large-4-1t-moe","2026-10-06T23:10:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"9fa3427c-cf69-46c7-9720-cd3b646a155b","B站开源35B翻译模型:3B激活,150种语言","bilibili-index-translate-35b-moe","2026-10-04T13:30:00+00:00"]