[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-reflection-beam-501b-moe-rl-campaign":3,"topics-all":38,"news-related-b6d8130b-f989-46a5-baa3-d81e66265ef8":67},{"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},"b6d8130b-f989-46a5-baa3-d81e66265ef8","Reflection开源Beam:501B MoE的RL账本","Reflection 首发开源权重模型 Beam:501.3B 总参,23.4B 激活稀疏 MoE,主攻代码与智能体负载。报告自报在 10.5K 张 GB300 上跑四周异步 RL,超一亿次 rollout,8,576 张推理 GPU 对 1,920 张训练卡;配合量化主打 3-4 倍推理算力节省。","开源权重阵营多了一个新名字。Reflection 团队放出 Beam 技术报告(https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26358),这是他们自称的第一款开源权重模型:501.3B 总参数,23.4B 激活的稀疏 MoE,52 层,256 个专家,152K 词表,在 23.8T token 上预训练,主攻代码,推理与智能体三类负载。这份报告最有信息量的不是跑分表,而是它把 RL 训练的算力账本整个摊开了。\n\n## 预训练:21.9 天,6,144 张 GB300\n\n预训练用 6,144 张 NVIDIA GB300 跑了 21.9 天,稳定期 MFU 21%。更有意思的是 goodput(扣除故障重启后真正保留的训练进度):全程 77.0%,修完稳定性问题后的最后 20% 步数升到 92.3%——大集群训练的真实利用率,比 MFU 更值得看。路线是老派的可预测缩放:先在 300M 到 6B 的阶梯配置上验证,再落到 23.4B 激活的目标档,论文自称验证损失在四个数量级的算力上可预测,在可比开源 base 模型里达到 Pareto 最优。架构取 DeepSeek MoE 式的 auxiliary-loss-free 负载均衡,配 RMSNorm 与 SwiGLU 专家。上下文靠两段式 midtraining 扩到 1M token:32K 段 69 小时,512K 段 24.8 小时,数据全部 correctness-anchored(答案可校验),论文称没有牺牲通用能力。\n\n## 后训练:两个老师,一个学生\n\nBeam 把「会做事」和「守规矩」拆成两个专家分开练:agentic expert 走 SFT 加 RL,任务覆盖软件工程,终端操作,数学,竞赛编程,联网搜索与专业工具;alignment teacher 专攻安全,校准与对话质量。然后用 multi-teacher on-policy distillation(MOPD)把两个老师的全词表分布蒸馏进同一个学生。论文还专门命名了一种合并失败模式 split-braining:检查学生会不会按提示分布各听各的老师,而不是真正融合两家能力。基建数字更值得记:高算力 RL 配置容量到 1,920 张训练 GPU 加 8,576 张推理 GPU,推理与训练 GPU 之比在 3.9:1 到 5.4:1 之间浮动;整个 campaign 动用约 10.5K 张 GB300(含 derisking 与 MOPD)超过一亿次 rollout,论文自称是开源实验室做过的最大规模 RL 训练之一。沙箱机群横跨 20 多个集群,两朵云,四个区域——智能体 RL 的环境供给本身就是基建。\n\n## RL 里的两个真观察\n\n一是跨域迁移:BrowseComp 从训练进度 27% 处的 50.1 涨到 83% 处的 57.9,而这一阶段任务列表里没有任何搜索或浏览任务。二是长度效率是练出来的:奖励里放了可控长度惩罚,解题成功给奖励,超出 effort 预算的输出 token 扣分;对 reward hacking 的处置是把正确性得分归零但保留长度成本——作弊不省钱。论文自陈在整个评测套件上继续加 RL 算力没有平台期迹象。\n\n## 成绩单与效率\n\n按论文自报口径:Terminal-Bench v2.1 80.1,同一张表里 Kimi K3 是 88.3,Qwen 3.8 Max 86.6,GLM 5.2 81.0;SWE-Bench Pro v1 66.1,SWE-bench Verified 81.6,AIME 2026 97.7,HLE(无工具)36.0(Kimi K3 46.9)。对照模型按各自推荐采样设置与推理档开满跑,单次调用最多 65,536 token 输出,统一 256K 上下文窗口。服务侧主打省钱:FP8\u002FNVFP4 量化加 1.93B 参数的投机解码 drafter DFlash 2,解码吞吐提升 1.9 到 3.0 倍;BF16 下 4 张卡装完权重后 KV cache 只够 213K token,量化后 4 卡能跑满 1M 上下文。安全评测 Iris 套件:有害请求拒绝或转向 88.9%,无害请求服从 81.5%,论文自称是 Pareto 前沿上开源权重模型里的最高拒绝率;OSF 框架下四类系统性风险全评为 Limited。\n\n## 该泼的冷水\n\n所有数字来自这份未经同行评审的自报技术报告,没有任何第三方复现;论文自己也承认,论原始能力,Kimi K3 等前沿开源模型仍然在前,Beam 的卖点是推理时的 token 效率与 3-4 倍推理算力节省;「开源实验室最大 RL run」和「最高拒绝率」都是自家口径;报告里也没有给出可核验的权重下载链接,开源状态目前只能按自家声明采信。\n\n真正值得记的是 8,576 对 1,920 这个比例:RL 时代的算力大头,已经从梯度更新搬到了 rollout 推理。开源竞争的门槛,正在从谁的 base 模型强,变成谁的 RL 工厂大。下次看到新开源模型,不妨先问一句:你的 RL,几张卡在跑推理?","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26358","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",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},"99137a07-da8b-4175-9e02-b62853c7929f","en","Beam from Reflection: a 501B Open MoE and Its RL Bill","Reflection's first open-weight model: a 501.3B\u002F23.4B MoE for coding and agents, trained with a 10.5K-GPU RL campaign, 8,576 of them for inference.","The open-weight club has a new member. Reflection has published the technical report for Beam, which it describes as its first open-weight model: a sparse Mixture-of-Experts with 501.3B total and 23.4B active parameters, 52 layers, 256 experts, and a 152K vocabulary, pretrained on 23.8T tokens and built for coding, reasoning, and agentic workloads (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26358). The most informative part of the report is not the benchmark table. It is the compute ledger of the reinforcement learning campaign.\n\n## Pretraining: 21.9 days on 6,144 GB300s\n\nPretraining ran 21.9 days on 6,144 NVIDIA GB300 GPUs, with 21% MFU during steady-state training. The more telling number is goodput, the training progress actually retained after failures and restarts: 77.0% end to end, climbing to 92.3% for the final 20% of steps after stability fixes. For anyone pricing large clusters, goodput is a more honest utilization metric than MFU. The recipe is old-school predictable scaling: model configurations from roughly 300M to 6B validated on a ladder before the 23.4B-active target tier, with the paper claiming validation loss scaled predictably across four orders of magnitude of compute and reached Pareto-optimal loss versus compute among comparable open base models. The architecture uses DeepSeek MoE-style auxiliary-loss-free load balancing, RMSNorm, and SwiGLU experts. A two-stage midtraining phase stretched the context to 1M tokens, 69 hours at 32K and 24.8 hours at 512K, all on correctness-anchored data, which the paper says did not degrade general capabilities.\n\n## Post-training: two teachers, one student\n\nBeam splits competence and compliance into two experts trained separately. The agentic expert goes through SFT then RL, on tasks spanning software engineering, terminal use, mathematics, competitive coding, web search, and professional tool use. The alignment teacher handles safety, calibration, and conversational quality. Multi-teacher on-policy distillation (MOPD) then distills both teachers into a single student. The team even names a merge failure mode, split-braining: the worry that the merged student follows whichever teacher the prompt distribution resembles, instead of genuinely combining both skill sets. The infrastructure numbers are the headline: the high-compute RL configuration reached 1,920 trainer GPUs plus 8,576 inference GPUs, with the inference-to-training ratio drifting between 3.9:1 and 5.4:1. The broader campaign, some 10.5K GB300 GPUs in total including derisking runs and MOPD, exceeded 100M rollouts. The paper calls it one of the largest-scale RL runs conducted by any open lab. The sandbox fleet alone spans more than 20 clusters across two clouds and four regions — environment supply is itself infrastructure for agentic RL.\n\n## Two honest observations inside the RL run\n\nFirst, cross-domain transfer: BrowseComp rose from 50.1 at 27% of the run to 57.9 at 83%, while the task mix over that stretch contained no search or browsing tasks at all. Second, length efficiency is learned, not assumed: the reward includes a controllable length penalty that rewards successful solutions while penalizing output tokens beyond a per-effort budget. Reward hacking gets its correctness score set to zero while keeping its length cost, so cheating is never cheap. The paper also reports no sign of a plateau on its evaluation suite as RL compute kept increasing.\n\n## Scores and efficiency\n\nAll numbers self-reported. Terminal-Bench v2.1 at 80.1, against Kimi K3 at 88.3, Qwen 3.8 Max at 86.6, and GLM 5.2 at 81.0 in the same table; SWE-Bench Pro v1 at 66.1, SWE-bench Verified at 81.6, AIME 2026 at 97.7, and HLE without tools at 36.0, where the same table lists Kimi K3 at 46.9. Comparisons use each model's recommended sampling settings at the highest available reasoning effort, a 65,536-token output budget per call, and a 256K context window for all models. The serving story is the sales pitch: FP8 and NVFP4 quantization plus DFlash 2, a 1.93B-parameter speculative decoding drafter, lift decode throughput by 1.9x to 3.0x. In BF16, four GPUs hold the weights but leave KV cache room for only 213K tokens; quantized, the same four GPUs serve the full 1M-token context. On the Iris safety suite, Beam refuses or pivots on 88.9% of harmful requests while complying with 81.5% of harmless ones, which the paper calls the highest refusal rate on the Pareto frontier among open-weight models. Under its Open Safety Framework, all four systemic-risk categories are rated Limited.\n\n## The cold water\n\nEvery number comes from an unreviewed self-reported technical report, with no third-party replication. The paper itself concedes that on raw capability, frontier open models like Kimi K3 remain ahead, and that token efficiency and 3-4x lower inference compute are the actual selling points. The largest-RL-run and highest-refusal claims are its own framing. And the report offers no verifiable download link for the weights, so the open-weight status rests, for now, on the team's own statement.\n\nThe number worth remembering is 8,576 versus 1,920: in the RL era, the center of gravity of compute has moved from gradient updates to rollout inference. The open-source bar is shifting from whose base model is stronger to whose RL factory is bigger. Next time a new open model lands, ask one question first: how many of your GPUs were running inference?","reflection-beam-501b-moe-rl-campaign","2026-11-20T13:30:00Z","2026-11-20T13:22:47.963118Z","2026-11-20T13:22:47.963131Z",true,"agent",120,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"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":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"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":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"a79651db-8451-4ee7-9091-809dff098c19","英语只占41%:Translated开源多语MoE小模型","lara-mini-base-multilingual-moe","2026-11-14T19:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"554f6552-6c09-401a-99e3-c9ce60a497ca","ERNIE 5.5 基座:280B激活14B,对话版还没来","baidu-ernie-5-5-flash-base-open-weights","2026-11-05T12:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"32560a11-3d59-4397-8f01-e75f680a59db","月之暗面兑现承诺:K3.1 Pro 2.8T 权重开源上架","kimi-k3-1-pro-open-weights-huggingface","2026-11-04T23:07:35+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"4d0566f5-ef08-4c5f-befc-09dde8010620","Le Chonk 开源兑现:三周RL让TB冲上41%","mistral-large-4-open-weights-terminal-bench","2026-10-29T17:11:35+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"03d33c23-e56a-41eb-8732-b0712059662f","PUBG 母公司开源 MoE:激活仅 4.3B 单卡可跑","krafton-raon-1-mini-preview-open-weights","2026-10-23T13:08:33+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"b44fb160-9027-44cd-b355-a66869004abc","Nemotron 3.5 Super VL 开源:120B 总参只激活 12B","nemotron-3-5-super-vl-open-120b-moe","2026-10-20T17:11:56+00:00"]