Reflection releases Beam, its first open-weights model
Reflection's first model, Beam, scores 30 on Artificial Analysis's Intelligence Index, which ranks it second among US open-weights models but leaves it well behind the proprietary frontier.
Reflection has released Beam, its first model. Beam is an open-weights, text-only mixture model with 501B total parameters, 23B of them active, and a 1M-token context window. The figures below come from independent testing by Artificial Analysis.
Beam scores 30 on the Artificial Analysis Intelligence Index. Among US open-weights models, only Meta's Muse Spark Open (39) scores higher, and Beam is ahead of NVIDIA's Nemotron 3.5 Super VL (26). All three came out this week. Across all open-weights models, Beam ranks tenth, one point behind the UAE's K2 Horizon (31). For comparison, the top proprietary model on the index, GPT-6 Astra (max), scores 53.
Strengths. Beam uses about 26k output tokens per Intelligence Index task, roughly half of K2 Horizon's ~52k. It scores 48.5% on AutomationBench-AA, 11 points ahead of K2 Horizon, and 13.1% on CritPt. It completes 34% of tasks on the Artificial Analysis Cyber Index, the highest score Artificial Analysis has recorded for a US open-weights model.
Weaknesses. Beam scores 3.0% on Terminal-Bench 4.0. Run in Mirror, Reflection's open-source CLI harness, it scores 28 on the Coding Agent Index. That is 15 points behind the next setup and 40 behind the leaders, though it uses very few tokens. On AA-Omniscience, Beam answers 33% of questions correctly, but its hallucination rate is 47%, compared with 26% for K2 Horizon.
The source does not include details on pricing, the license or training.