Reflection AI releases Apache 2.0-licensed weights for Beam, a 501B-parameter model trained with RL to penalize unnecessary tokens
| Source: Techmeme | Original article
Reflection AI released Apache 2.0‑licensed weights for Beam, its first open 501‑billion‑parameter model trained with RL that penalizes unnecessary tokens, challenging Chinese open models.
Reflection AI announced on October 16 that it has made the weights of Beam, its first open‑weight model, publicly available under an Apache 2.0 license. Beam is a sparse mixture‑of‑experts (MoE) system with 501 billion total parameters but only 23 billion active at inference time. The company says the model was trained with reinforcement learning that penalises unnecessary token generation, aiming to finish coding, reasoning and agentic tasks more efficiently.
The release marks a notable shift in the open‑model landscape. By publishing the full weight set on Hugging Face and partnering with Baseten to host the model on its platform, Reflection AI gives developers unrestricted commercial and modification rights—something still rare for models of this scale. The token‑efficiency focus lets Beam achieve reasoning performance comparable to GLM 5.2 while using three to four times less compute, though it lags behind leading Chinese offerings such as Kimi K3 on agentic coding benchmarks.
The timing is significant. Just a day earlier, Meta unveiled Muse Spark Open, a 816 billion‑parameter model that also carries an open‑weight licence. As we reported on October 15, Meta’s entry raised the bar for open‑weight scale, and Reflection’s Beam provides a more compute‑lean alternative that could appeal to startups and research groups with limited resources. The move also intensifies competition with Chinese open models that have dominated the high‑parameter segment.
What to watch next includes community adoption of Beam’s weights, real‑world benchmark results beyond the initial comparisons, and whether Reflection AI will follow up with larger or specialised variants. The broader industry will also monitor how the Apache 2.0 licence influences commercial deployment and whether other firms adopt similar token‑efficiency training regimes to curb inference costs.
Sources
Back to AIPULSEN