Reflection AI unveils 501-billion-parameter open-weight model Beam to rival Chinese options
Nvidia-backed startup Reflection AI has unveiled Beam, a 501-billion-parameter open-weight model designed for coding, reasoning, and AI-agent tasks to compete with leading Chinese open models.
Reflection AI, an artificial intelligence startup founded by two former Google DeepMind researchers and backed by Nvidia, has unveiled Beam, its first AI model. The company positions the open-weight release as a challenger to leading open-weight options from China, asserting that Beam excels at coding, reasoning, and agentic workflows while using three to four times less inference compute than comparable models.[8][9][5][7][3]
Beam is built on a sparse mixture-of-experts architecture with 501 billion total parameters, 23 billion active parameters, and a 1-million-token context window. According to technical specifications reported by Wall St Engine, the base model was pretrained on 23.8 trillion tokens using 6,144 Nvidia GB300 GPUs in under four weeks. Its reinforcement-learning run engaged 10,500 Nvidia GB300 GPUs over four weeks, generating more than 100 million rollouts and utilizing 46.4 million sandboxes per day. To support its infrastructure, Reflection has signed over $7 billion in compute agreements with SpaceX and Nebius through 2029.[3][10][1][6]
The company plans to release Beam's weights, model card, technical report, and evaluation tools under the Apache 2.0 license later this month. Ahead of the release, Reflection shared benchmark scorecards indicating mixed outcomes against competitors: while Reflection reports that Beam scored 80.9 on SWE-Bench Verified and rivals GLM-5.2 on reasoning, it trailed models such as Qwen 3.8-Max on Terminal-Bench v2.1 and other benchmarks.[3][10][2][4]
Key facts
- Reflection AI, founded by two former Google DeepMind researchers and backed by Nvidia, unveiled Beam as its first open-weight AI model.
- Beam features a sparse mixture-of-experts architecture with 501 billion total parameters, 23 billion active parameters, and a 1-million-token context window.
- Pretraining was completed on 23.8 trillion tokens using 6,144 Nvidia GB300 GPUs in under four weeks.
- The reinforcement-learning training ran for four weeks across 10,500 Nvidia GB300 GPUs, generating over 100 million rollouts and utilizing 46.4 million sandboxes per day.
- Reflection AI has signed over $7 billion in compute agreements with SpaceX and Nebius to secure GB300 infrastructure through 2029.
- Reflection AI scheduled Beam's model weights, technical report, and model card for release under the Apache 2.0 license in October 2026.
- Reflection claims Beam performs comparably to GLM-5.2 on reasoning using 3x to 4x less inference compute, though the benchmark claims have not been independently verified.
Sources · 10 sources
- WC
Wccftech@wccftechPost on X ·
NVIDIA-backed Reflection used 10,500x GB300 GPUs and 46.4 million sandboxes per day for 4 weeks to train Beam, a 501B open-weight model that is insanely efficient. 🔗 https://t.co/VzPdetkwIF https://t.co/8JJKJAQrva
Open source - MT
MTS@MTSlivePost on X ·
SITUATION DETECTED: Reflection AI announced Beam, an open-weight model that advances the Western open frontier on coding and agentic tasks. Weights are due later this month. https://t.co/uJRALICFx4
Open source - WS
Wall St Engine@wallstenginePost on X ·
Nvidia-backed Reflection AI has launched Beam, a 501B-parameter open-weight AI model built for coding, reasoning and AI-agent workloads. Beam is aimed at competing with leading open-weight models from China. Beam is a text-only mixture-of-experts model with: • 501B total parameters, 23B active • 1M-token context window • Pretrained on 23.8T tokens • Built for coding, reasoning, tool use and agentic workloads Reflection says Beam performs comparably with Z .ai’s GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute. The company’s benchmark claims have not yet been independently verified. Beam’s reinforcement-learning run used 10,500 NVIDIA $NVDA GB300 GPUs for four weeks and generated more than 100M rollouts. Its base model was pretrained on 6,144 GB300 GPUs in under four weeks. Reflection is targeting enterprises, developers and sovereign governments, with a broader strategy around customized “AI factories” that can connect organizations’ proprietary data to locally controlled AI systems. The NVIDIA-backed startup has also signed more than $7B worth of compute deals with SpaceX and Nebius $NBIS to secure access to GB300 infrastructure through 2029. Reflection plans to release Beam’s weights under an Apache 2.0 license later this month, along with its technical report, model card and tools for running, evaluating and fine-tuning the model.
Open source - R�
RuntimeWire 🏴☠️@runtimewirePost on X ·
Reflection published Beam's benchmark scorecard ahead of releasing weights; results are mixed. Beam is 501B parameters, trained on 23.8T tokens. Terminal-Bench 2.1: Beam 80.1 vs GLM 5.3 88.2. https://t.co/1na7SKwQld Source: Financial Times https://t.co/MxzDH7c4rC
Open source - TE
TechCrunch@TechCrunchPost on X ·
Nvidia-backed Reflection AI unveils Beam, its first open-weight model, which it says rivals GLM-5.2 on reasoning with far less inference compute. Weights are due this month. https://t.co/YhJvdSQpTE
Open source - SB
Shay Boloor@StockSavvyShayPost on X ·
Reflection just launched Beam which is a 501B open weight model for coding, reasoning and agents that it says can match GLM-5.2 while using ~4x less inference compute. The more interesting part is RL used 10,500 $NVDA GB300s versus 6,144 for pretraining and generated 100M+ rollouts showing agent training is becoming a huge inference workload. That helps explain why Reflection locked in $7B+ of compute across $SPCX and $NBIS through 2029 while positioning Beam as a U.S. open model for enterprises and sovereign AI.
Open source - TE
Techmeme@TechmemePost on X ·
Reflection unveils open-weight model Beam, saying it excels at coding and agentic tasks, uses 3x-4x less compute than comparable models, and nears Qwen 3.8-Max (Semafor) (Visit Techmeme dot com for the link and full context!)
Open source - BL
Bloomberg@businessPost on X ·
Reflection AI, an artificial intelligence startup from two former Google DeepMind researchers, has unveiled a new open-weight model that it says rivals leading options in the US and China https://t.co/8agHGvBLW4
Open source - RE
Reuters@ReutersPost on X ·
Nvidia-backed Reflection unveils first AI model to take on Chinese open models https://t.co/RuxWnnp6YR https://t.co/RuxWnnp6YR
Open source - 阳明
阳明AI@x_autonomyPost on X ·
【AI热点】02:00-04:00 更多详细信息→https://t.co/8waYU02Yi5 2. Epoch AI 报告：OpenAI 研究员 coding-agent 推理支出约每月翻倍 [影响大·可执行低]Epoch AI 报告显示，OpenAI 研究员的 coding-agent 推理支出按 API 牌价计算约每月翻倍。到 2026 年 8 月中，中位数研究员每天花费 $601、一年约 $2M，大致持 3. Reflection 发布 501B 开源权重模型 Beam [影响大·可执行低]Reflection 推出首个开源权重模型 Beam，为稀疏 MoE 架构，总参数 501B、激活 23B，主打编码、推理和智能体任务，权重、技术报告和模型卡将于本月以 Apache 2.0 许可发布 4. Understanding AI 分析智能体集群为何可能成为下一个 scaling law [影响小·可执行中]Understanding AI 撰文分析多智能体协作是否构成新的 scaling law。文章指出 OpenAI 通过让模型在多智能体环境中训练并用消息工具互通，实现了数千个智能体协作，如 1000 5. Meta 和 Microsoft 大幅削减内部 Claude 使用 [影响大·可执行低]据 The Information，Meta 和 Microsoft 两大企业客户大幅削减内部使用 Claude。 6. Claude Code 推出 mods，可自定义其界面、提示词与工具调用 [影响大·可执行低]Claude Code 现已支持 mods，即用户编写的函数，可改变 Claude Code 的渲染内容、进入提示词的内容以及工具调用。Anthropic 的 Lydia 展示了三个示例：实时上下文分 7. 弗吉尼亚州长 Spanberger 发布新能源规划，让 AI 数据中心增长服务于零碳目标 [影响大·可执行低]弗吉尼亚州民主党州长 Abigail Spanberger 于 10 月 1 日发布四年一度的州能源规划，坚持 Virginia Clean Economy Act 中世纪零碳目标，并提出四条到 20 8. TikTok 推出 AI 购物助手和一键结账功能 [影响大·可执行低]TikTok 于周一宣布推出对话式 AI 购物助手和应用内一键结账功能，用户可在 For You 信息流中直接向品牌购买。该助手能理解上下文、记住用户偏好，并提供商品详情、物流、尺码、库存等实时购物指 9. Instinct 将 AI 智能体引入群聊，好友无需注册账号也能用 [影响中·可执行高]估值 100 亿美元的 Instinct 宣布，用户可将 AI 智能体加入群聊，用于旅行规划、抢票、组队梦幻联赛、拼车等场景，好友即使未注册 Instinct 也能参与。该功能今日起向早期访问用户开放 10. Reka AI 发布 19B 全能模型 Rho-1，单模型处理文本、图像、视频与机器人控制 [影响大·可执行低]Reka AI 发布 190 亿参数全能模型 Rho-1 的研究预览版，在单一神经网络内处理并生成文本、图像、视频和机器人控制动作。该模型将全部模态作为 token 放入同一上下文窗口，无需工具调用或 11. Agent Arena 智能体能力榜：Anthropic 与 OpenAI 领跑 [影响大·可执行低]Agent Arena 按解决跨领域问题的智能体能力对模型排名，Anthropic 模型在 Code、Work、Chat 三个领域均居第一，OpenAI 的 GPT 6 Astra (Max) 三项全 12. llama.cpp v0.6.0 发布 [影响大·可执行低]全新的 v0.6.0 版本带来了许多好东西： - 支持 Clef（文本 + 视觉）- 对 Qwen3.8-Flash-Next 的高质量支持- Metal 性能大幅提升- 新增 `llama_ 13. 独立的未来在于相互依存 [影响小·可执行中]Angie Dixon 在文章中提出，独立的未来在于相互依存，而非凡事亲力亲为。她以自身经历说明，可调节床、抓取工具、送货服务、远程问诊等辅助手段并未妨碍独立，反而让她的生活得以运转。她指出，独立意味 14. JEPA-Anything：跨世界学习预测模型 [影响中·可执行高]JEPA-Anything： 跨不同世界学习预测模型 15. Claude Code 的 HTML 计划技能 [影响大·可执行低]我一直在做一个技能，让 Claude Code 生成更好的 HTML 计划。 它使用简单的语言，展示代码片段，呈现问题并制作 mockup。Linting 减少了 Claude 常遇到的典型失败情况 16. Skyfall AI 用世界模型预测企业决策后果 [影响中·可执行高]由 Maluuba 团队创立的 Skyfall AI 正在构建预测企业每次决策后如何变化的世界模型。其创始人认为，LLM 过去 5 年主要靠增加数据、算力和模型规模进步，但长周期规划能力差、数据需求巨 17. Claude Code 插件市场安装体验 [影响大·可执行低]安装后请通过以下方式给我反馈： claude plugin marketplace add anthropics/claude-plugins-community claude plugin in 18. Google Cloud 数据库如何为生产级 AI 激活数据层 [影响中·可执行高]Google Cloud 发布系列实验，演示如何用 AlloyDB 和 Cloud SQL 为生产级 AI 准备数据：在数据库中生成文本嵌入做语义搜索，为 RAG 提供 grounding，并支持多模 19. Codex 0.160.1 修复 Windows 远程 MCP 环境变量保留问题 [影响中·可执行高]Codex 发布 0.160.1，修复了在显式配置远程环境变量启动远程 stdio MCP 服务器时 SYSTEMROOT、TEMP 和 TMP 被丢弃的问题，使 Unix 主机能够保留 Window 20. 分类器三种类型：二分类、多分类与多标签 [影响小·可执行中]分类器把输入映射到固定标签集合，分为二分类（邮件∈{垃圾、非垃圾}）、多分类（文本∈{正面、中性、负面}）和多标签（电影⊆{动作、恐怖、喜剧、爱情、奇幻、惊悚}）三类。其工作方式是先用逻辑回归等手工特 21. Understanding AI 迎来新作者 Dan Kagan-Kans [影响小·可执行中]Understanding AI 宣布新作者 Dan Kagan-Kans 加入，他此前做了十年 Mosaic 杂志执行主编，过去一年以自由撰稿人身份报道 AI，曾撰写《左派正在错失 AI》、在 Ne 22. Thariq 分享代码调用栈与标注功能 [影响小·可执行中]我特别喜欢调用栈（这个是从 @dillon_mulroy 那看到的）以及给代码片段加标注的能力 23. Viggle Animate 一键重现梵高 [影响中·可执行高]梵高依然是最伟大的作曲家。毋庸置疑。 用 Viggle Animate，一键用你喜欢的角色重现它。🎨 链接 ↓ 24. ARC Prize 2026 峰会主讲嘉宾公布 [影响小·可执行中]公布 ARC Prize Research Summit 2026 主讲嘉宾 Kenneth Stanley @kenneth0stanley 是神经进化和开放式 AI 领域的先驱，NEAT 和新 25. 苹果 visionOS 27.2 开发者预览版 Beta 3 发布 [影响小·可执行中]苹果向 Vision Pro 用户推送 visionOS 27.2 开发者预览版 Beta 3，内部版本号 24N5103f，距上次 Beta 发布间隔 14 天。因各区域节点服务器缓存问题，部分地区 26. Suno 推荐 Oakwood///diskrot 新专辑 naenia [影响小·可执行中] http 27. Dream Relic 发布新专辑《Lost In a Dream》 [影响小·可执行中]继 Seven-Eleven Halo 与 Time is a Limited Allowance 走红之后，Dream Relic 发布了他的新专辑《Lost In a Dream》。 https 28. Suno 上 ECHLO 新专辑发布 [影响小·可执行中]ECHLO 的新项目既有电影感又令人萦绕于心。这张专辑有着精美呈现的钢琴片段，围绕强劲推进的节奏舞动。一切都由空灵的人声支撑，初次聆听后仍久久萦绕。 https://t.co/vUDlUilTLc 29. ARC Prize 2026 研究峰会开放报名 [影响小·可执行中]申请参会 https://t.co/yhEzSN94lv 30. Suno 发布专辑 Old Soul [影响小·可执行中]Old Soul 是一张关于变老过程中成长烦恼的专辑。在温暖的木吉他伴奏下，Isaiah Wallace 回望常青树、甜茶和夏末夜晚。 https://t.co/1Jw2OhZ5Td 31. GPU 为何售价 6000 美元 [影响小·可执行中]不错 😀 （引用推文：这就是 GPU 卖到 6000 美元的原因） 32. Suno 推出 KakerMix 80 年代复古风专辑 [影响小·可执行中]受 80 年代标志性声音启发，KakerMix 使用复古合成器、有力的电子鼓和空灵混响，让我们仿佛进入另一个维度。 https://t.co/4QXAUsJXlx 33. Mistral CEO 发文表兴奋 [影响小·可执行中]兴奋
Open source