Reflection AI has introduced Beam, its first open-weight sparse Mixture-of-Experts (MoE) model designed for coding, reasoning, and agentic workloads. The model features 501B total parameters with only 23B active per token, aiming to compete with larger open models while using significantly less inference compute.
- Beam was pretrained on 23.8 trillion tokens using 6,144 NVIDIA GB300 GPUs in under four weeks, with a curated dataset that retained high-quality samples conventional filters would drop.
- Reinforcement learning utilized 10.5K NVIDIA GB300 GPUs for four weeks, generating over 100 million rollouts across nearly one million coding and agentic environments.
- The model achieves 80.9 on SWE-bench Verified and 80.1 on Terminal Bench v2.1, outperforming Nemotron 3 Ultra on the former while remaining close to GLM 5.2 on the latter.
- Beam includes a controllable reasoning effort parameter, allowing users to adjust answer length and complexity based on task difficulty and compute budget.
- Weights are scheduled for release under an Apache 2.0 license later in October 2026, with early access currently available via a waitlist on the Reflection platform.
Reflection positions Beam as advancing the Western open-weight frontier by offering efficiency at inference time compared to competitors like Kimi K3 and GLM 5.2.