跳到主内容
@wquguru
精选88Hacker News Best(web_list)模型发布/更新多源精选 ×4

Reflection发布501B开源模型Beam,主打编码与智能体高效推理

Beam: Reflection's 501B open-weight model

原文
发到 X
推荐理由

全新开源MoE大模型,在编码与智能体场景下展现出显著的推理效率优势,技术细节详实,值得关注。

We are introducing Beam, Reflection’s first open-weight model. Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active, built for coding, reasoning, and agentic workloads.

我们在此介绍 Beam,Reflection 的首个开放权重模型。Beam 是一个稀疏混合专家(Mixture-of-Experts)模型,总参数量为 5010 亿,激活参数量为 230 亿,专为编码、推理和智能体工作负载而构建。

Beam’s capabilities come from major investments in both pretraining and reinforcement learning (RL). We pretrained the model on 23.8 trillion diverse, curated, high-quality tokens from the web and proprietary licensed datasets, matching or outperforming available similar-sized open base models. In parallel, we developed the algorithms, training environments, and infrastructure needed to sustain high-compute RL at exceptional scale. Our high-compute RL run generated over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over 4 weeks of training.

Beam 的能力源于我们在预训练和强化学习(RL)方面的重大投入。我们在来自网络和专有授权数据集的 23.8 万亿个多样化、精心策划的高质量 token 上对模型进行了预训练,其表现达到或超越了现有的类似规模的开放基础模型。同时,我们开发了维持超大规模高算力强化学习所需的算法、训练环境和基础设施。我们的强化学习运行在为期 4 周的训练中,使用了 10500 块 NVIDIA GB300 GPU,生成了超过 1 亿次 rollout。

Together, these efforts produced competitive open-weight performance with frontier inference compute efficiency.

通过这些努力,我们实现了具有竞争力的开放权重性能,并具备前沿推理计算效率。

Beam is undergoing final red-teaming and evaluations. You can sign up here for early access to the model. We will release the weights, technical report, model card, and developer artifacts later this month.

Beam 正在进行最后的红队测试和评估。您可以在此注册以获取模型的早期访问权限。我们将在本月晚些时候发布权重、技术报告、模型卡片和开发者工具。

Model Capability

模型能力

We trained Beam with a particular focus on coding and agentic performance. Beam advances the Western open-weight frontier and is competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks. Where frontier open models like Kimi K3 remain ahead on raw capability, Beam's advantage is efficiency at inference time.

我们训练 Beam 时特别关注编码和智能体性能。Beam 推动了西方开放权重的前沿,在与 GLM 5.2 等更大规模的开放模型竞争中表现出色,并在编码和智能体任务上接近 Qwen 3.8-Max。虽然像 Kimi K3 这样的前沿开放模型在原始能力上仍领先,但 Beam 的优势在于推理时的效率。

The below figure shows Beam's performance across a range of coding, agentic, reasoning, and STEM benchmarks. NR denotes scores that have not been reported.

下图展示了 Beam 在一系列编码、智能体、推理和 STEM 基准测试中的表现。NR 表示未报告的分数。

Agentic Coding/TerminalReasoningTool Calling / Search General Capabilities

智能体编码/终端推理工具调用/搜索通用能力

Row labelBeamInklingNemotron 3 UltraGLM 5.2GLM 5.3Kimi K3Qwen 3.8 MaxDeepSeek V4.1 Flash
DeepSWE v1.144.4NRNR44.061.068.051.074.2
SWE Bench Pro v2-Hard77.256.9NRNR84.388.2NRNR
SWE Bench Pro v165.554.346.462.1NRNR67.7NR
Terminal Bench v2.180.163.856.481.088.288.386.690.6
SWE Atlas Codebase QnA34.6NRNRNR61.068.0NRNR
SWEBench Multilingual78.0NR67.7NRNRNRNRNR
SWEBench Verified80.977.670.7NRNRNRNRNR
行标签BeamInklingNemotron 3 UltraGLM 5.2GLM 5.3Kimi K3Qwen 3.8 MaxDeepSeek V4.1 Flash
DeepSWE v1.144.4NRNR44.061.068.051.074.2
SWE Bench Pro v2-Hard77.256.9NRNR84.388.2NRNR
SWE Bench Pro v165.554.346.462.1NRNR67.7NR
Terminal Bench v2.180.163.856.481.088.288.386.690.6
SWE Atlas Codebase QnA34.6NRNRNR61.068.0NRNR
SWEBench Multilingual78.0NR67.7NRNRNRNRNR
SWEBench Verified80.977.670.7NRNRNRNRNR
Row labelBeamInklingNemotron 3 UltraGLM 5.2GLM 5.3Kimi K3Qwen 3.8 MaxDeepSeek V4.1 Flash
AIME 202697.897.1NR99.2NRNRNRNR
HLE no tools36.229.726.740.542.346.943.639.1
SciCode49.746.144.6NR59.058.752.152.0
CriPT AA16.35.43.120.919.123.420.014.3
GPQA Diamond90.587.28791.291.793.592.690.9
行标签BeamInklingNemotron 3 UltraGLM 5.2GLM 5.3Kimi K3Qwen 3.8 MaxDeepSeek V4.1 Flash
AIME 202697.897.1NR99.2NRNRNRNR
HLE no tools36.229.726.740.542.346.943.639.1
SciCode49.746.144.6NR59.058.752.152.0
CriPT AA16.35.43.120.919.123.420.014.3
GPQA Diamond90.587.28791.291.793.592.690.9
Row labelBeamInklingNemotron 3 UltraGLM 5.2GLM 5.3Kimi K3Qwen 3.8 MaxDeepSeek V4.1 Flash
AutomationBench public37.0NRNR26.248.246.739.854.8
MCP Atlas78.776.063.177.884.282.384.5NR
tau3 banking38.025.022.637.1NR37.155.2NR
BrowseComp w/ context management77.477.144.4NRNR91.2NRNR
DeepSearchQA w/context management80.1NRNRNRNR95.0NRNR
行标签BeamInklingNemotron 3 UltraGLM 5.2GLM 5.3Kimi K3Qwen 3.8 MaxDeepSeek V4.1 Flash
AutomationBench public37.0NRNR26.248.246.739.854.8
MCP Atlas78.776.063.177.884.282.384.5NR
tau3 banking38.025.022.637.1NR37.155.2NR
BrowseComp w/ context management77.477.144.4NRNR91.2NRNR
DeepSearchQA w/context management80.1NRNRNRNR95.0NRNR
Row labelBeamInklingNemotron 3 UltraGLM 5.2GLM 5.3Kimi K3Qwen 3.8 MaxDeepSeek V4.1 Flash
AA-LCR79.377.379.378.379.788.780.384.0
LongBench v265.5NR61.964.0NRNR66.3NR
IFBench79.779.881.773.3NRNR82.8NR
AA Omniscience (index) - public split only (our results)13.014.28.6NR22.6NRNR6.6
行标签BeamInklingNemotron 3 UltraGLM 5.2GLM 5.3Kimi K3Qwen 3.8 MaxDeepSeek V4.1 Flash
AA-LCR79.377.379.378.379.788.780.384.0
LongBench v265.5NR61.964.0NRNR66.3NR
IFBench79.779.881.773.3NRNR82.8NR
AA Omniscience (index) - public split only (our results)13.014.28.6NR22.6NRNR6.6

Beam pairs coding and agentic capabilities with highly efficient reasoning. On advanced reasoning benchmarks, it achieves scores comparable to GLM-5.2 while using 3–4× less inference compute. Efficiency gains are even more pronounced when comparing to models in the 2T+ parameter family like Qwen 3.8-Max, which require significantly more inference compute per token.

Beam 具备高效的代码与智能体能力,以及高度优化的推理性能。在高级推理基准测试中,其得分可与 GLM-5.2 相媲美,同时推理计算量仅为后者的 3–4 倍。与 Qwen 3.8-Max 等参数量在 2T+ 的模型相比,效率优势更为显著,因为后者每个 token 所需的推理计算量大得多。

These results translate into more intelligence per token, delivering strong model capabilities at lower cost, making Beam a powerful workhorse model for enterprise coding and agentic workloads.

这些结果意味着每个 token 能带来更高的智能水平,以更低成本提供强大的模型能力,使 Beam 成为企业级代码编写和智能体工作负载的强大主力模型。

Figure 2: Beam demonstrates frontier-level inference efficiency, both when measured in terms of FLOPS and token count across DeepSWE, Humanity’s Last Exam (HLE), and Terminal Bench 2.1. We used data from Artificial Analysis and DataCurve, estimating generation forward-pass compute as FLOPs ≈ 2 × active parameter count × mean generated tokens per attempt, counting each multiply-add as two operations. Generated tokens include both reasoning and the final answer. For mixture-of-experts models, we used the parameters activated per token rather than the total model size. These estimates exclude prompt prefill, context-dependent attention operations, and serving overhead, so they represent an approximate compute comparison rather than measured inference cost. We use Artificial Analysis and DataCurve as sources for other model’s evals.

图2:Beam 展示了前沿级别的推理效率,无论是在 DeepSWE、Humanity’s Last Exam (HLE) 和 Terminal Bench 2.1 上以 FLOPS 还是 token 数量衡量。我们使用了 Artificial Analysis 和 DataCurve 的数据,将生成分前向传播计算量估算为 FLOPs ≈ 2 × 激活参数量 × 每次尝试生成的平均 token 数,并将每次乘加运算计为两次操作。生成的 token 包括推理过程和最终答案。对于混合专家(MoE)模型,我们使用每个 token 激活的参数而非模型总参数量。这些估算排除了提示词预填充(prompt prefill)、上下文相关的注意力操作以及服务开销,因此它们代表的是近似的计算量比较,而非实测的推理成本。我们使用 Artificial Analysis 和 DataCurve 作为其他模型评估数据的来源。

High-Compute Reinforcement Learning

高计算强化学习

We made high-compute reinforcement learning a central scaling axis for Beam, investing in RL science, data, and infrastructure to turn more compute into stronger capabilities. Scaling RL enables more extensive exploration of problem-solving strategies, while longer rollouts support multi-step reasoning, tool use, and adaptation to environment feedback.

我们将高计算强化学习作为 Beam 的核心扩展轴,在强化学习科学、数据和基础设施方面进行投入,旨在将更多的计算资源转化为更强的能力。扩展强化学习能够更广泛地探索问题解决策略,而更长的 rollout(轨迹)则支持多步推理、工具使用以及对环境反馈的适应。

To scale reinforcement learning, we deployed 10.5K NVIDIA GB300 GPUs for four weeks generating more than 100 million rollouts with a maximum context length of 256K tokens. Training and grading used approximately 1.3 billion sandboxes. To sustain a run of this magnitude, we sourced one million high-quality coding, agentic, and STEM environments. We believe this is one of the largest scale RL runs conducted by any open lab to date. Across our evaluation suite, capabilities continued to improve as we increased RL compute, with no sign of a plateau.

为了扩展强化学习,我们部署了 10,500 块 NVIDIA GB300 GPU,持续四周,生成了超过 1 亿条 rollout,最大上下文长度为 256K tokens。训练和评分使用了约 13 亿个沙箱。为了维持如此规模的运行,我们获取了一百万个高质量编码、智能体(agentic)和 STEM 环境。我们认为这是迄今为止任何开放实验室进行的规模最大的 RL 运行之一。在我们的评估套件中,随着 RL 计算量的增加,能力持续提升,未见 plateau(平台期)迹象。

Figure 3: Terminal-Bench 2.1, HLE, and DeepSWE scores as a function of cumulative RL rollouts during training of Beam’s reasoning expert, accounting for 80M of the over 100M rollouts generated across the full RL campaign. For comparison, Inkling was trained on 30M rollouts and MiMo on 753K.

图3:Beam 推理专家训练期间,Terminal-Bench 2.1、HLE 和 DeepSWE 得分随累积 RL rollout 数量的变化函数,其中包含了整个 RL 活动中生成的超过 1 亿条 rollout 中的 8000 万条。作为对比,Inkling 在 3000 万条 rollout 上训练,MiMo 在 75.3 万条 rollout 上训练。

We trained Beam with asynchronous policy gradients. At scale, policy staleness becomes a major source of instability for these methods. Long running rollouts have tokens that are generated by multiple model checkpoints, with earlier tokens becoming increasingly stale relative to the current policy. Numerical mismatch between training and inference engines further compounds this challenge.

我们使用异步策略梯度对 Beam 进行训练。在大规模下,策略陈旧性成为此类方法不稳定的主要来源。长时间运行的 rollout 包含由多个模型检查点生成的 token,较早生成的 token 相对于当前策略变得日益陈旧。训练引擎与推理引擎之间的数值不匹配进一步加剧了这一挑战。

We developed new algorithms to maintain stable learning under these conditions while systematically reducing training–inference mismatch throughout our pipeline. These advances enable fully asynchronous RL at scale that remains stable, even when learning from interactions generated more than a day earlier.

我们开发了新算法,在这些条件下保持稳定的学习过程,并系统地减少整个流程中的训练-推理不匹配。这些进展实现了大规模完全异步强化学习(RL),即使在从一天前生成的交互中学习时,也能保持稳定。

Figure 4: Stable learning continues as staleness builds up over time. The top plot shows the oldest sample in the batch, with the bottom plot demonstrating stable numerics. Even when training Beam with one-day staleness—107 weight versions behind the current policy—the numerics remain stable.

图4:随着时间推移陈旧度增加,稳定学习仍在继续。上图显示了批次中最旧的样本,下图展示了稳定的数值表现。即使在使用一天陈旧度的数据训练Beam时——即权重版本比当前策略落后107个版本——数值依然保持稳定。

Learning to reason efficiently

高效推理学习

We trained Beam with a controllable length penalty that rewards successful solutions while discouraging unnecessary tokens. Early in RL, performance improved even as completion lengths fell: the model learned to solve tasks more effectively with less reasoning. Later, as Beam developed stronger agentic capabilities, completion lengths grew again, but those additional tokens supported further gains in performance. Throughout training, RL improved the tradeoff between capability and token usage.

我们使用可控的长度惩罚对Beam进行训练,该惩罚奖励成功的解决方案,同时抑制不必要的token生成。在强化学习早期,尽管完成长度下降,性能仍有所提升:模型学会了用更少的推理步骤更有效地解决问题。后期,随着Beam发展出更强的代理能力,完成长度再次增长,但这些额外的token支持了性能的进一步提升。在整个训练过程中,强化学习改善了能力与token使用之间的权衡。

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 →