Reflection发布开源模型Beam,SemiAnalysis评测Clau
[AINews] Reflection Beam - 501B-A23B American Open Model
涵盖美国新开源模型Beam的技术细节与定位,以及Claude对OpenAI订阅价值的深度对比评测,信息密度高且具行业风向标意义。
It’s been over a year since Reflection launched with us with big goals on coding (and hinted about their RL approach):
Reflection 带着宏大的编程目标(并暗示了其强化学习方法)与我们合作已有一年多的时间:
But they stayed “stealth” longer than Thinking Machines and it was not clear we would ever get a model launch out of them, as the broader open model ecosystem did not slow down one bigt for them. Well, we did:
但他们保持“隐形”的时间比 Thinking Machines 更长,我们也不清楚他们能否推出模型,因为更广泛的开源模型生态并没有为他们放慢脚步。好吧,我们确实做到了:
They compare themselves to Inkling, Nemotron, and GLM 5.2, but the SOTA GLM 5.3, Kimi K3, Qwen 3.8 Max, and DeepSeek V4.1 Flash are generally ahead. Still, since this is US-trained from scratch, there’s a segment of the market that has been eagerly waiting for more options here, and more importantly, Reflection has now announced its arrival as a functional neolab!
他们将自身与 Inkling、Nemotron 和 GLM 5.2 进行比较,但 SOTA 的 GLM 5.3、Kimi K3、Qwen 3.8 Max 和 DeepSeek V4.1 Flash 通常更胜一筹。尽管如此,由于这是从头开始在美国训练的,市场上有一部分人一直在热切等待更多的选择,更重要的是,Reflection 现在宣布其作为功能性新实验室的到来!
AI News for 10/03/2026-10/5/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!
2026年10月3日至10月5日的AI新闻。我们检查了12个 subreddit、544条推文,没有进一步的 Discord 信息。AINews 的网站允许你搜索所有过往期号。提醒一下,AINews 现在是 Latent Space 的一个板块。你可以选择加入或退出电子邮件频率!
AI Twitter Recap
AI Twitter 回顾
Reflection’s Beam Leads a Wave of Open-Weight Releases
Reflection 的 Beam 引领了一波开放权重发布的浪潮
- Beam launch: Reflection announced Beam, a text-only 501B-total / 23B-active MoE for coding, agentic and scientific work. It was trained from scratch, and full weights under Apache 2.0 are due this month (announcement, Laskin).
- Training scale: Team posts cite 23.8T pretraining tokens, partly from an OCR pipeline over hundreds of millions of PDFs (data lead). They also describe a stable RL/OPD run on 10K GB300s with more than 100M rollouts across ~1M tasks (Damos).
- Claimed results: A summary of Reflection’s claims gives 80.9 on SWE-bench Verified, 3–4x the inference efficiency of GLM 5.2, and four weeks each of pretraining and RL on ~10,500 GB300s (summary). A tech report and OSS integrations are promised (Polozov).
- Context: Axios reported the launch ahead of time. It said Reflection pays $150M/month for Colossus compute plus a $1B Nebius deal, and that other unnamed US labs will ship open models this month (Curran).
- Independent and critical reads: Artificial Analysis has early access and expects Beam to be among the most token-efficient open models for its intelligence (AA).
- MFU and architecture: Elie Bakouch estimates only ~12% BF16 MFU in pretraining. He reads the architecture as 3:1 interleaved global/sliding-window attention and notes better held-out code perplexity than DSv4 (analysis).
- Compute comparison: Teortaxes calls Beam an iso-FLOP replication of DeepSeek V3 (post). He infers ~1.3B RL sandboxes over 4 weeks, with up to 170K running at once (sandboxes).
- Positioning: Observers place Beam around GLM-5.2 level (iScienceLuvr) and below DSv4 Flash on some benchmarks (critique). Nathan Lambert groups it with Nvidia and Thinking Machines as strong US releases that still trail Chinese counterparts (Lambert).
- Other open and specialized models:
- Aleph Alpha Kolibri: 78B total / 3.46B active, Apache 2.0, built for German and English. Self-reported scores are 96.9% AIME 2025, 84.3% GPQA Diamond and 66.4% SWE-Bench Verified (summary). The dataset is unreleased, and agentic evals sit well below Qwen (Jitsev).
- Reka Rho-1: A 19B omni model that understands and generates text, images, video and robot actions, trained from scratch on 320 H100s in ~3 months (announcement, compute).
- Decision models: Command Code’s Agr (31B) and Agr-flash (360M) skip text generation and return typed values with per-option probabilities for tool calls and routing (Agr). SemiAnalysis explains that TypeSafe’s Jev uses the same no-decode approach and displaces frontier models mainly in router roles (explainer).
- Smaller releases: Upstage’s Solar Mini 4 (35B / 3B active, 512K context) is free on Nous Portal for two weeks (Nous). Eleven v4 Turbo tops AA’s Provider Voice TTS arena at half the price of v4 (AA).
- Beam 发布:Reflection 宣布了 Beam,这是一个仅用于文本的501B总参数/23B活跃参数的 MoE 模型,适用于编程、智能体和科学工作。它是从头训练的,Apache 2.0 下的完整权重将于本月发布(公告,Laskin)。
- 训练规模:团队帖子引用了23.8T预训练token,部分来自对数亿PDF的OCR管道(数据负责人)。他们还描述了在10K GB300上进行的稳定RL/OPD运行,跨越约1M任务进行了超过100M次 rollout(Damos)。
- 声称的结果:Reflection 声称的结果摘要显示,SWE-bench Verified 得分为80.9,推理效率是 GLM 5.2 的3-4倍,并且在约10,500个GB300上分别进行了四周的预训练和RL(摘要)。承诺提供技术报告和OSS集成(Polozov)。
- 背景:Axios 提前报道了此次发布。它表示 Reflection 每月为 Colossus 计算支付1.5亿美元,并有一笔10亿美元的 Nebius 交易,其他未具名的美国实验室将在本月推出开源模型(Curran)。
- 独立和批判性阅读:Artificial Analysis 拥有早期访问权,预计 Beam 将成为其智能方面最具 token 效率的开源模型之一(AA)。
- MFU 和架构:Elie Bakouch 估计预训练中的 BF16 MFU 仅为约12%。他将架构解读为3:1交错的全球/滑动窗口注意力,并指出其保留的代码困惑度优于 DSv4(分析)。
- 计算比较:Teortaxes 称 Beam 是对 DeepSeek V3 的等 FLOP 复制(帖子)。他推断在四周内有约1.3B个RL沙箱,同时最多有170K个在运行(沙箱)。
- 定位:观察者将其置于 GLM-5.2 级别(iScienceLuvr)并在部分基准测试中低于 DSv4 Flash(批评)。Nathan Lambert 将其与 Nvidia 和 Thinking Machines 归为一类,认为它们是美国发布的强劲模型,但仍落后于中国同类产品(Lambert)。
- 其他开源及专用模型:
- Aleph Alpha Kolibri:总参数量 78B / 激活参数 3.46B,Apache 2.0 协议,专为德语和英语构建。其自报得分为 AIME 2025 达 96.9%、GPQA Diamond 达 84.3%、SWE-Bench Verified 达 66.4%(摘要)。该数据集尚未发布,且智能体评估表现远低于 Qwen(Jitsev)。
- Reka Rho-1:一款 19B 全能模型,能理解和生成文本、图像、视频及机器人动作,在约 3 个月内使用 320 块 H100 GPU 从零训练而成(公告,算力)。
- 决策模型:Command Code 的 Agr(31B)和 Agr-flash(360M)跳过文本生成,直接返回带有选项概率的键入值,用于工具调用和路由(Agr)。SemiAnalysis 解释称,TypeSafe 的 Jev 采用相同的无解码方法,主要在路由器角色中取代前沿模型(解析)。
- 较小规模发布:Upstage 的 Solar Mini 4(35B / 3B 激活,512K 上下文)在 Nous Portal 上免费开放两周(Nous)。Eleven v4 Turbo 以 v4 一半的价格在 AA 的 Provider Voice TTS 竞技场中登顶(AA)。
OpenAI vs Anthropic: Subscription Value, Speed and Evals
OpenAI 与 Anthropic:订阅价值、速度与评估
- SemiAnalysis limit testing: SemiAnalysis tested plans from Anthropic, OpenAI, Meta, SpaceXAI, MiniMax, Moonshot, Cursor, Cognition and others. It found that Claude subscriptions deliver 5x+ more API-equivalent value than OpenAI plans (report).
- Methodology: Value depends on the credit cost of each model and token type, not on list API prices (thread).
- Task-cost adjustment: Adjusting for task cost narrows Claude’s edge to 1.3–2.9x (scaling01).
- Unverified compute estimate: One analyst claims Anthropic spends 42% of inference compute on subscriptions that earn ~10% of revenue (chart).
- OpenAI capacity squeeze: Users report that new $200 sign-ups were paused and that usage limits were effectively halved across plans. GPT-6.1 Sol was positioned as the efficient alternative (analysis). Theo describes a reversal in coding-model preference between July and September (post).
- OpenAI response: Codex lead Tibo pledged a meaningful improvement or a full reset every day for 28 days (pledge).
- Day 1 speedup: Default speed for GPT-6 Astra and GPT-6.1 Sol rose ~50%, from ~30 to ~50 TPS. The change covers all subscription surfaces and Sign in with ChatGPT partners such as OpenCode, Pi, Amp and Devin (day 1, TPS).
- Friction: Banked Codex resets expire without timezone adjustment (report). The always-on dots agent is limited to $100+ Pro plans (criticism).
- Enterprise demand (reported): The Information reports that Microsoft cut projected internal Anthropic spend by more than a third. It also reports Meta’s Claude Code users fell from ~60K to ~30K, largely because of a push to Meta’s own tools (summary).
- Leaderboards:
- Agent Arena: Anthropic holds #1 in Code, Work and Chat. Fable 5.1 leads Code and Work, while GPT-6 Astra places #2 in Code (Arena).
- Design Arena: GPT-6 Astra is #1 in 3D Design, Frontend, Full Stack and Image-to-HTML (Design Arena).
- Hallucination: On AA-Omniscience, Gemini 4 Argon guesses wrong on 15% of questions it doesn’t know, versus 29% for the next best model. GPT-6 Astra has the highest accuracy at 61% (data).
- SemiAnalysis 极限测试:SemiAnalysis 测试了来自 Anthropic、OpenAI、Meta、SpaceXAI、MiniMax、Moonshot、Cursor、Cognition 等公司的计划。发现 Claude 订阅提供的 API 等效价值是 OpenAI 计划的 5 倍以上(报告)。
- 方法论:价值取决于每种模型和令牌类型的信用成本,而非列表 API 价格(线程)。
- 任务成本调整:考虑任务成本后,Claude 的优势缩小至 1.3–2.9 倍(scaling01)。
- 未经证实的算力估算:一位分析师声称 Anthropic 将 42% 的推理算力用于订阅,这些订阅仅贡献约 10% 的收入(图表)。
- OpenAI 产能挤压:用户报告称新注册 $200 套餐已暂停,且各套餐的使用限制实际上减半。GPT-6.1 Sol 被定位为高效替代方案(分析)。Theo 描述了 7 月至 9 月期间对编码模型偏好的逆转(帖子)。
- OpenAI 回应:Codex 负责人 Tibo 承诺在 28 天内每天实现显著改进或全面重置(承诺)。
- 第 1 天提速:GPT-6 Astra 和 GPT-6.1 Sol 的默认速度提升约 50%,从约 30 TPS 增至约 50 TPS。此变更涵盖所有订阅界面以及通过 ChatGPT 登录的合作方,如 OpenCode、Pi、Amp 和 Devin(第 1 天,TPS)。
- 摩擦:Banked Codex 重置过期时未进行时区调整(报告)。始终在线的 dots agent 仅限 $100+ Pro 套餐使用(批评)。
- 企业需求(报道):The Information 报道称,微软将内部 Anthropic 支出预期削减了三分之一以上。该报道还指出,Meta 的 Claude Code 用户从约 6 万降至约 3 万,主要原因是推动使用 Meta 自有工具(摘要)。
- 排行榜:
- Agent Arena:Anthropic 在代码、工作和聊天类别中排名第一。Fable 5.1 在代码和工作类别中领先,GPT-6 Astra 在代码类别中排名第二(Arena)。
- Design Arena:GPT-6 Astra 在 3D 设计、前端、全栈和图像转 HTML 类别中排名第一(Design Arena)。
- 幻觉:在 AA-Omniscience 上,Gemini 4 Argon 对不知晓的问题有 15% 答错,而次优模型的错误率为 29%。GPT-6 Astra 准确率最高,达 61%(数据)。
Agent Harnesses, RL Environments and Developer Tooling
Agent Harnesses、RL 环境与开发者工具
- Multi-harness RL (Hugging Face): A capture proxy speaks the OpenAI Chat, OpenAI Responses, Anthropic and Gemini formats. It forwards calls to vLLM and records exact token IDs and logprobs for TRL, so 10 unmodified harnesses become RL environments (Delangue, explainer).
- Results: The same weights score 62% under Mini-SWE-Agent and 33% under Claude Code. Training LFM2.5-2.6B across 4 harnesses lifts first-attempt solves from 42% to 54%, and a tool-call bonus cuts calls by 31%. SFT on 3,189 rollouts plateaus at 47.5%.
- Caveats: The run used one task family and one seed.
- Environment hosting: RL environments are now hosted and versioned on the HF Hub like datasets (blog).
- Pi Durable: Earendil’s harness is built around a small task-based workflow engine, so long-running, multiplayer agents can suspend and resume anywhere (Pi).
- Design: The core is ~15K lines of TypeScript with SQLite/JSONL storage and runs on Bun or Cloudflare Durable Objects. Control is separated from execution environments (review).
- Effect.ts: The authors explain they skipped Effect because it does not provide durability (Zechner).
- Agent memory: Cognition launched Devin “Dreaming,” which prunes and links a memory graph overnight. It is open-sourcing the git- and markdown-backed format as Agent Memory Repo (launch, format).
- Cursor SDK: The update adds mid-run steering, background subagents that report back to the parent, replaceable system prompts, and MCP readOnlyHint and destructiveHint annotations on custom tools (steering, annotations).
- DeepSeek Harness: An experimental Claude Code Mods compatibility layer in v0.2.1-alpha.1 tests whether DSH’s “everything is a plugin” architecture is a superset of Claude Code’s extension points (team post).
- Routing and access:
- Cline: Its Pareto 26.10 Preview routes across models and grades answers, claiming $0.24 versus $13.41 per task at equal DeepSWE score (Cline). Cline also paused its free DeepSeek-V4.1-Flash promotion over abuse (notice).
- 多 harness RL(Hugging Face):一个捕获代理支持 OpenAI Chat、OpenAI Responses、Anthropic 和 Gemini 格式。它将调用转发至 vLLM,并记录精确的 token ID 和 logprobs 以供 TRL 使用,从而使 10 个未经修改的 harness 成为 RL 环境(Delangue,解释)。
- 结果:同一权重在 Mini-SWE-Agent 下得分为 62%,在 Claude Code 下为 33%。在 4 个 harness 上训练 LFM2.5-2.6B 模型,首次尝试解决率从 42% 提升至 54%,工具调用奖励使调用次数减少 31%。在 3,189 次 rollout 上进行 SFT 后,性能在 47.5% 处趋于平稳。
- 注意事项:本次运行仅使用了一个任务家族和一个随机种子。
- 环境托管:RL 环境现在像数据集一样在 HF Hub 上进行托管和版本控制(博客)。
- Pi Durable:Earendil 的 harness 围绕一个小型基于任务的流程引擎构建,因此长时间运行的多人 agent 可以在任意位置挂起和恢复(Pi)。
- 设计:核心部分约为 1.5 万行 TypeScript 代码,使用 SQLite/JSONL 存储,可在 Bun 或 Cloudflare Durable Objects 上运行。控制逻辑与执行环境分离(评论)。
- Effect.ts:作者解释说他们跳过了 Effect,因为它不提供持久性(Zechner)。
- Agent 记忆:Cognition 推出了 Devin “Dreaming”,它能在 overnight 期间修剪和链接记忆图。它正在开源基于 git 和 markdown 的记忆格式,称为 Agent Memory Repo(发布,格式)。
- Cursor SDK:此次更新增加了运行中引导、向父级汇报的后端子 agent、可替换的系统提示词,以及自定义工具上的 MCP readOnlyHint 和 destructiveHint 注解(引导,注解)。
- DeepSeek Harness:v0.2.1-alpha.1 版本中的一项实验性 Claude Code Mods 兼容层,旨在测试 DSH 的“一切皆插件”架构是否为 Claude Code 扩展点的超集(团队博文)。
- 路由与访问:
- Cline:其 Pareto 26.10 Preview 版本可在不同模型间进行路由并对答案进行评级,声称在 DeepSWE 得分相同的情况下,每任务成本为 0.24 美元,而对比对象为 13.41 美元(Cline)。由于遭到滥用,Cline 还暂停了其免费提供的 DeepSeek-V4.1-Flash 推广活动(公告)。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力