Reflection AI Announces Beam: 501B MoE with Open Weights

Reflection AI has introduced Beam, its first open-weights model: a sparse MoE with 501 billion parameters (23 billion active) and a 1 million token context, achieving the best result among open models on SWE-Bench Verified (80.9), with a promise to release the weights under Apache 2.0 this month.

Compiled by Sergey KostenchukPublished 2026-10-06Updated 2026-10-06

2026-10-06 Coding NVIDIA

💻 Reflection AI Announces Beam: 501B MoE with Open Weights

The lab's first open model: a sparse MoE with 501 billion parameters (23 billion active), pre-trained on 23.8 trillion tokens, with an effective context of up to 1 million. Best among open models on SWE-Bench Verified (80.9), but inferior to Kimi K3 in raw power. The claimed "3–4 times fewer FLOPs" than GLM 5.2 is partly a result of the counting methodology: without prefill, attention, and serving overhead.

🌍 The real value is the RL infrastructure: over 100 million rollouts in 4 weeks on 10,500 NVIDIA GB300s, asynchronous policy gradients, robust to weight lag of 107 versions. A claim to be an open Western alternative to Kimi K3 and Qwen.

👤 Weights under Apache 2.0 and a technical report are promised "this month"; early access is already open. There are no comparisons on an equal compute budget — conclusions about efficiency are premature.

Source 1: https://reflection.ai/blog/introducing-beam Source 2: https://platform.reflection.ai/

Sources