Beam：Reflection 的 501B 开源权重模型（还不能下载） Beam: Reflection’s 501B open-weight model (not downloadable yet)

来源Sources

为什么推荐Why we recommend it 博文写明：仍在最终 red-teaming（安全对抗测试）和评测中；权重、技术报告、模型卡和 developer artifacts 这个月晚些时候才发布，许可写的是 Apache 2.0。所以现在不能当已可下载的开源模型。表格数字、3–4 倍推理计算量、23.8 万亿 token、GPU 数量、1 亿次尝试、256K / 1M 上下文都是厂商自述，本库未复跑、未下权重。与 gemini-4-argon、gpt-6-astra、claude-opus-5-5 互链。 The static HTML shows no weight download, and this library did not sign up. The post says final red-teaming and evaluations are still underway, and that the weights, technical report, model card, and developer artifacts come later this month under Apache 2.0. It is not a downloadable open model today. The table, the 3–4× inference-compute claim, 23.8 trillion tokens, the GPU counts, the 100 million rollouts, and the 256K / 1M context figures are vendor claims — not re-run here, weights not downloaded. Featured, unrated. Cross-linked with gemini-4-argon, gpt-6-astra, and claude-opus-5-5.

要点（官方称 / 社区称）Highlights (vendor / community claims)

• 官方称：稀疏混合专家，总参数 5010 亿、每次启用 230 亿，面向写代码、推理和让助手干活。模型本身只处理文字；图片或别的材料要先变成文字它才能用。现在还不能下载。博文写正在做最后的安全测试，提前试用只是登记；权重、技术报告、模型卡和开发者配套「这个月晚些时候」放出，许可打算用 Apache 2.0。 Vendor claim: a sparse mixture-of-experts model, 501 billion parameters total and 23 billion active, built for coding, reasoning, and agent work. The model itself is text-only; images or other material have to be turned into text first. It is not downloadable yet. The post says final safety testing is still underway and early access is a signup. Weights, the technical report, the model card, and developer extras are promised later this month, under Apache 2.0.
• 官方表格（本库未复跑，NR 表示对方没公布）：Beam 一列里 DeepSWE v1.1 为 44.4，Terminal Bench v2.1 为 80.1，SWE Bench Pro v1 为 65.5，SWEBench Verified 为 80.9，不用工具的 HLE 为 36.2，GPQA Diamond 为 90.5。正文另称：在更难的推理测试上，分数和 GLM-5.2 差不多，推理计算量大约少 3 到 4 倍。这是厂商比较，不是本库测出来的名次。 Vendor table (not re-run here; NR means the other lab did not publish a score). In the Beam column: DeepSWE v1.1 44.4, Terminal Bench v2.1 80.1, SWE Bench Pro v1 65.5, SWEBench Verified 80.9, HLE with no tools 36.2, GPQA Diamond 90.5. The prose also says that on harder reasoning tests the scores are comparable to GLM-5.2 at about 3–4× less inference compute. That comparison is the vendor’s, not a ranking produced here.
• 官方称：预训练用了 23.8 万亿 token（文本切成的小段），在 6144 张 NVIDIA GB300 NVL72 上不到四周跑完。强化学习用 1.05 万张 NVIDIA GB300，四周里产生超过 1 亿次尝试，单次上下文最长 256K token；文中写中期训练把可用上下文拉到 100 万 token。这些都是厂商自述，本库没有复现训练。 Vendor claim: pretraining used 23.8 trillion tokens and finished in under four weeks on 6,144 NVIDIA GB300 NVL72 GPUs. Reinforcement learning used 10,500 NVIDIA GB300 GPUs, produced more than 100 million rollouts over four weeks, and allowed up to 256K tokens of context; the post says midtraining extends usable context to 1 million tokens. All of that is the vendor’s account. This library did not reproduce the training.