The details
A mixture-of-experts model splits its network into many specialist parts and uses only a few for each token, so it runs faster than its size suggests. Beam still needs the memory of its full size: Hugging Face suggests planning for four Nvidia B200 GPUs to serve its compressed 4-bit version. It ships with its own Rust-based harness, plus recipes for fine-tuning and reinforcement learning. On Reflection’s own figures, it scores 81.6% on SWE-bench Verified, a coding test. Reflection is offering $10,000 in Hub credits for the best projects built with it.
Why it matters
Most frontier-scale models are rented through APIs. An openly licensed model of this size lets companies keep code and data in-house and tailor the model to their own work. The catch is cost: running it still takes several top-end GPUs. The benchmark scores come from Reflection itself, so independent tests will show whether Beam holds up against closed coding models.