跳到主内容
@wquguru
精选88MarkTechPost(RSS)模型发布/更新多源精选 ×4

Reflection AI发布Beam:501B参数MoE模型

Reflection AI Introduces Beam: A 501B Open-Weight MoE Model With 23B Active Parameters for Coding and Agentic Workloads

原文
发到 X
推荐理由

全新开源MoE模型发布,架构细节与RL工程实践披露详尽,对关注推理效率与Agent能力的开发者有直接参考价值。

Reflection AI has introduced Beam, its first open-weight model. Beam is a sparse Mixture-of-Experts (MoE) model with 501B total parameters and 23B active per token, built for coding, reasoning and agentic workloads. As per the Reflection AI team, Beam directly competes with larger open models like GLM 5.2 while using 3 to 4x less inference compute on reasoning benchmarks.

Reflection AI 推出了 Beam,这是其首个开放权重模型。Beam 是一个稀疏混合专家(MoE)模型,总参数量为 501B,每个 token 激活 23B,专为编码、推理和智能体工作负载而构建。据 Reflection AI 团队介绍,Beam 在推理基准测试中使用的计算量比 GLM 5.2 等更大的开源模型少 3 到 4 倍,从而直接与之竞争。

Is it deployable today? Not for self-hosting yet. Beam is in final red-teaming. Early access runs through a waitlist on the Reflection platform.

现在可以部署吗?目前还不能自托管。Beam 正处于最终的红队测试阶段。早期访问需要通过 Reflection 平台上的等待名单进行。

What is Reflection Beam?

什么是 Reflection Beam?

Beam is a general agent model trained from scratch by Reflection AI. It targets enterprise coding and agentic workloads. Reflection positions Beam as advancing the Western open-weight frontier. The research team is candid about the gap. Kimi K3 stays ahead on raw capability, so Beam’s pitch is efficiency at inference time.

Beam 是由 Reflection AI 从头训练的通用智能体模型。它针对企业级编码和智能体工作负载。Reflection 将 Beam 定位为推动西方开放权重前沿的技术。研究团队坦诚地指出了差距:Kimi K3 在原始能力上仍占领先,因此 Beam 的卖点在于推理时的高效性。

Users get a reasoning effort parameter. Lower settings favor short answers. Higher settings allow longer reasoning on hard tasks. Teams can match effort to task difficulty and compute budget.

用户可以获得一个推理努力参数。较低的设置倾向于简短的回答。较高的设置允许在困难任务上进行更长的推理。团队可以根据任务难度和计算预算来匹配努力程度。

How Beam was Pretrained

Beam 的预训练过程

Beam was pretrained on 23.8 trillion tokens from the web, public sources and proprietary licensed datasets. Reflection team states that its curation removed about 95% of raw internet tokens. It also kept roughly 1.8 trillion high-quality tokens that conventional filters would have dropped.

Beam 在来自互联网、公共来源和专有授权数据集的 23.8 万亿个 token 上进行预训练。Reflection 团队表示,其数据筛选过程移除了约 95% 的原始互联网 token。它还保留了大约 1.8 万亿个高质量 token,这些 token 通常会被常规过滤器丢弃。

The architecture interleaves local and global attention with fine-grained routed experts. Load balancing builds on auxiliary-loss-free balancing from DeepSeek-V3, adding cosine decay of expert-bias updates. The busiest expert reached just 1.04x average load at the end of pretraining. Across all 52 layers, residual norms stayed bounded using depth-based scaling, SandwichNorm, attention gating and FP32 residual accumulation.

该架构将局部注意力和全局注意力与细粒度的路由专家交错排列。负载均衡基于 DeepSeek-V3 的无辅助损失平衡方法,并增加了专家偏差更新的余弦衰减。在预训练结束时,最繁忙的专家负载仅为平均负载的 1.04 倍。在所有 52 层中,通过使用基于深度的缩放、SandwichNorm、注意力门控和 FP32 残差累积,残差范数保持在有界范围内。

Pretraining finished in under 4 weeks on 6,144 NVIDIA GB300 NVL72 GPUs. Goodput reached 92.3% near the end, with 9 semi-automatic rewinds. Midtraining extended effective context to 1M tokens.

预训练在 6,144 块 NVIDIA GB300 NVL72 GPU 上不到 4 周内完成。接近结束时,有效吞吐量达到 92.3%,进行了 9 次半自动回滚。中期训练将有效上下文扩展至 100 万 token。

High-Compute Reinforcement Learning

高计算强化学习

RL is Beam’s central scaling axis. The run used 10.5K NVIDIA GB300 GPUs for 4 weeks and generated over 100 million rollouts. Maximum rollout context was 256K tokens. Training and grading consumed about 1.3 billion sandboxes across nearly 1 million coding, agentic and STEM environments.

强化学习是 Beam 的核心扩展轴。此次运行使用了 10,500 块 NVIDIA GB300 GPU,持续 4 周,生成了超过 1 亿次 rollout。最大 rollout 上下文为 256K token。训练和评分消耗了约 13 亿个沙箱环境,涵盖近 100 万个编码、智能体和 STEM 环境。

Reflection trained with fully asynchronous policy gradients. Every token is tagged with the policy version that produced it. New algorithms kept learning stable even at one-day staleness, 107 weight versions behind the current policy. The team reports no plateau as RL compute increased.

采用完全异步策略梯度进行训练的反思机制。每个标记都带有生成它的策略版本标签。新算法即使在一天延迟、落后当前策略107个权重版本的情况下,仍能保持学习稳定。团队报告称,随着强化学习计算量的增加,性能没有出现停滞。

Infrastructure numbers are notable. The system sustained 110K concurrent rollouts on average. New weights reached the inference fleet in a median of about 12 seconds. 71 inference incidents were handled without stopping training.

基础设施数据引人注目。该系统平均维持了11万次并发 rollout( rollout 指从初始状态到终止状态的完整轨迹)。新权重到达推理集群的中位时间约为12秒。处理了71次推理故障,且未中断训练。

A controllable length penalty taught Beam to solve tasks with fewer tokens. Browsing skills also improved without browsing tasks in the RL mix, which suggests transfer across agentic domains.

一种可控的长度惩罚机制教会 Beam 使用更少的标记完成任务。即使在不包含浏览任务的强化学习混合训练中,其浏览技能也得到了提升,这表明在智能体领域之间存在迁移能力。

Safety and Alignment

安全与对齐

Reflection trained a separate safety and alignment teacher from the pretrained checkpoint. It merged that teacher with the RL teacher using multi-teacher on-policy distillation. Safety training used deliberative alignment. Safety evaluation results will appear in the technical report.

反思机制从预训练检查点训练了一个独立的安全与对齐教师模型。它通过多教师在线蒸馏将该教师模型与强化学习教师模型合并。安全训练采用了深思熟虑的对齐方法。安全评估结果将出现在技术报告中。

Benchmarks (Reflection-Reported)

基准测试(Reflection 报告)

On SWE-bench Verified, Beam scores 80.9 versus 70.7 for Nemotron 3 Ultra. On Terminal Bench v2.1, Beam scores 80.1, close to GLM 5.2 at 81.0. DeepSeek V4.1 Flash (90.6) and Kimi K3 (88.3) lead there. These numbers come from Reflection’s table, which sources rival scores from Artificial Analysis and DataCurve.

在 SWE-bench Verified 上,Beam 得分为80.9,而 Nemotron 3 Ultra 为70.7。在 Terminal Bench v2.1 上,Beam 得分为80.1,接近 GLM 5.2 的81.0。DeepSeek V4.1 Flash(90.6)和 Kimi K3(88.3)在该榜单上领先。这些数据来自 Reflection 的表格,其中竞争对手的分数来源于 Artificial Analysis 和 DataCurve。

Beam vs Closest Open-Weight Competitors

Beam 与最接近的开源权重竞争对手对比

FeatureReflection BeamGLM-5.2Nemotron 3 UltraDeepSeek V4.1 FlashKimi K3
DeveloperReflection AI (US)Z.ai (China)NVIDIA (US)DeepSeek (China)Moonshot AI (China)
Total params501B~753B550B552B backbone + 196B Engram2.8T
Active params23B~40B55B8B prefill / 16B decode104B
Context1M (effective)1MUp to 1M1M1M
InputTextTextTextText + imageText + image
LicenseApache 2.0 (planned)MITOpenMDW-1.1MITKimi K3 License
WeightsLater in Oct 2026AvailableAvailableAvailableAvailable
Terminal Bench v2.1*80.181.056.490.688.3
SourceReflectionHugging FaceNVIDIAHugging FaceHugging Face
特性Reflection BeamGLM-5.2Nemotron 3 UltraDeepSeek V4.1 FlashKimi K3
开发者Reflection AI (美国)Z.ai (中国)NVIDIA (美国)DeepSeek (中国)Moonshot AI (中国)
总参数量501B~753B550B552B 主干 + 196B Engram2.8T
激活参数量23B~40B55B8B 预填充 / 16B 解码104B
上下文窗口1M (有效)1M最高 1M1M1M
输入类型文本文本文本文本 + 图像文本 + 图像
许可证Apache 2.0 (计划中)MITOpenMDW-1.1MITKimi K3 许可证
权重发布2026年10月下旬已提供已提供已提供已提供
Terminal Bench v2.1*80.181.056.490.688.3
数据来源ReflectionHugging FaceNVIDIAHugging FaceHugging Face

*Scores as published in Reflection’s Beam announcement. Specs verified October 5, 2026.

*分数发布于 Reflection 的 Beam 公告中。规格于2026年10月5日核实。

Beam is the smallest model here by total parameters. Its 23B active count sits below GLM-5.2, Nemotron 3 Ultra and Kimi K3. Apache 2.0 and MIT are standard permissive licenses. Kimi K3’s custom license adds attribution requirements for very large products.

Beam 是此处总参数量最小的模型。其23B的激活参数量低于 GLM-5.2、Nemotron 3 Ultra 和 Kimi K3。Apache 2.0 和 MIT 是标准的宽松许可证。Kimi K3 的自定义许可证对超大型产品增加了署名要求。

Key Takeaways

关键要点

  • Beam is a 501B MoE with 23B active parameters, text-only, and 1M effective context.
  • Pretraining used 23.8T tokens on 6,144 NVIDIA GB300 GPUs in under 4 weeks.
  • RL ran 100M+ rollouts on 10.5K GB300 GPUs over 4 weeks.
  • Reflection reports 80.9 on SWE-bench Verified and 80.1 on Terminal Bench v2.1.
  • Apache 2.0 weights are scheduled for later this month.
  • Beam 是一个拥有 23B 活跃参数的 501B MoE 模型,仅支持文本,有效上下文长度为 1M。
  • 预训练在不到 4 周的时间内,使用 6,144 块 NVIDIA GB300 GPU 处理了 23.8T tokens。
  • 强化学习在 4 周内于 10.5K 块 GB300 GPU 上运行了超过 1 亿次 rollout。
  • Reflection 在 SWE-bench Verified 上得分为 80.9,在 Terminal Bench v2.1 上得分为 80.1。
  • Apache 2.0 许可证的权重预计将于本月晚些时候发布。

Check out the Technical details and Early Access.

查看技术细节和早期访问权限。

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

进入量化体系 →