Details
- Reflection AI released Beam 1.0, its first open-weight model, with an NVFP4 checkpoint available through Hugging Face.
- Beam uses a sparse mixture-of-experts architecture with 501 billion total parameters and 23 billion active parameters per token.
- NVIDIA said Beam was trained on GB300 NVL72 systems and delivers more than 70,000 tokens per second per GPU on Blackwell Ultra out of the box.
- The model targets coding, reasoning, and agentic workloads, where selective expert activation can reduce compute per token compared with dense models of similar scale.
- Reflection trained Beam on 6,144 GB300 NVL72 GPUs in under four weeks and used a separate 10,500-GPU reinforcement-learning run that generated more than 100 million rollouts.
- The release expands access from early testing to deployable model weights, with support emerging across tools including Hugging Face Transformers, vLLM, SGLang, and NVIDIA Dynamo.
- Beam is documented with a context window of up to one million tokens, although real-world performance will depend on hardware, quantization, serving software, and workload.
Impact
Beam gives the open-model ecosystem a large, agent-focused alternative optimized for NVIDIA’s newest infrastructure rather than only a research preview. Its sparse design and NVFP4 release could make a 501B-parameter model more practical to serve, while the claimed Blackwell performance strengthens NVIDIA’s hardware-software case. The launch also raises competitive pressure on other open-model developers to pair large parameter counts with lower inference costs and immediate deployment support.