Reflection has published the weights for Beam, its first model, making a 501-billion-parameter mixture-of-experts system available under the Apache 2.0 license.
Only about 23 billion parameters activate per token, and the model accepts text with a context window of one million tokens. It targets coding and agentic work, and ships with a dedicated open-source harness written in Rust.
Trained on a large GPU fleet
Reflection pretrained Beam on 23.8 trillion tokens, then refined it through a reinforcement learning run that generated more than 100 million rollouts on 10,500 Nvidia GB300 GPUs over four weeks.
The company first showed the model on October 5 while it was still undergoing red-teaming, with early access limited to a waitlist. It released weights in BF16, FP8 and NVFP4 formats plus a speculative decoding drafter, and published a safety technical report alongside the model card.
Day-zero support arrived across Transformers, vLLM and SGLang, and the company listed more than 40 partners and integrations. Dell and HPE will deliver Beam behind the firewall for on-premises and air-gapped deployments, and a public beta of Reflection’s own API is open and free to start.
An operating system pitch
Reflection describes the arrangement as an open operating system for agents, with the model as the kernel and integrations layered above the infrastructure. It is also soliciting adoption directly: the best projects built on Beam will split $10,000 in Hugging Face credits.
The release continues a run of US open-weight launches this week, alongside Meta’s Muse Spark Open and Nvidia’s Nemotron additions. Apache 2.0 terms make Beam the most permissively licensed of the group, which matters to companies that need to modify and redistribute what they host.
