Reflection AI发布开源模型Beam,性能对标GLM-5.2
The much anticipated open-source model from Reflection AI just dropped.
The much anticipated open-source model from Reflection AI just dropped.
Reflection AI 备受期待的开源模型刚刚发布。
Beam now becomes the strongest western open model, rivals GLM-5.2 while using 3-4x less inference compute.
Beam 现已成为最强的西方开源模型,可与 GLM-5.2 抗衡,同时推理计算量仅为后者的 3-4 倍。
> The 3-4x efficiency edge over GLM-5.2 rests on estimated forward-pass FLOPs (2 × active parameters × generated tokens), which leave out prefill, attention and serving overhead, so it is not a measured cost.
> Beam 相比 GLM-5.2 拥有 3-4 倍的效率优势,这一优势基于估算的前向传播 FLOPs(2 × 活跃参数量 × 生成的 token 数),未包含预填充、注意力机制和服务开销,因此并非实测成本。
> A sparse mixture-of-experts model with 501B total and 23B active parameters, pretrained on 23.8T tokens in under 4 weeks on 6,144 GB300 GPUs, with a 1M-token effective context.
> 这是一个稀疏混合专家(MoE)模型,总参数量为 501B,活跃参数量为 23B,在不到 4 周的时间内使用 6,144 块 GB300 GPU 在 23.8T 个 token 上完成预训练,有效上下文长度为 1M token。
> The RL run used 10.5K GB300 GPUs for 4 weeks, produced over 100M rollouts across nearly 1M environments and about 1.3B sandboxes, and ended with no sign of a plateau.
> 强化学习(RL)运行使用了 10.5K 块 GB300 GPU,历时 4 周,在接近 1M 个环境和约 1.3B 个沙箱中产生了超过 1 亿次 rollout,且未见收敛迹象。
> Beam nearly ties GLM-5.2 at 80.1 versus 81.0 on Terminal Bench v2.1, but trails DeepSeek V4.1 Flash at 90.6 and Kimi K3 at 88.3, neither of which appears in the headline chart.
> 在 Terminal Bench v2.1 上,Beam 的得分为 80.1,几乎与 GLM-5.2 的 81.0 持平,但落后于 DeepSeek V4.1 Flash 的 90.6 和 Kimi K3 的 88.3,这两者均未出现在主图表中。
and Reflection will also ship FP8 and NVFP4 builds under Apache 2.0.
此外,Reflection 还将以 Apache 2.0 许可证提供 FP8 和 NVFP4 构建版本。
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