Reflection Releases Beam, an Open 501B Model, Claims 81.6% on SWE-Bench Verified
Summary
Reflection releases Beam, a 501B-parameter open mixture-of-experts model that activates 23B parameters per token and handles 1M-token contexts under Apache 2.0. The company reports 81.6% on SWE-bench Verified at BF16 precision and offers $10,000 in Hugging Face Hub credits for the best projects built on it.
Key Points
- Reflection's Beam routes 8 of 256 experts per layer across 52 decoder layers.
- The NVFP4 checkpoint cuts Beam's weights from 1 TB in BF16 to roughly 280 GB for Blackwell GPUs.
- Hugging Face says the hybrid attention cache grows 36 KiB per token, about a sixth of the full-attention rate.