Reflection announces Beam, a 501B open-weight coding model

Reflection announced Beam, a 501B open-weight MoE with 23B active parameters, trailing some open coding leaders while claiming lower inference compute.

Reflection AI announced Beam, a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active per token, built for coding, reasoning, and agentic work. Apache 2.0 weights are planned later this month after final red-teaming; The Decoder adds that an early version is already available to select users. Reported results include 80.9 on SWE-bench Verified, 80.1 on Terminal-Bench v2.1, and 44.4 on DeepSWE v1.1. Help Net Security says the Terminal-Bench score trails GLM 5.2 (81.0), Kimi K3 (88.3), and DeepSeek V4.1 Flash (90.6), while Reflection claims performance comparable to GLM 5.2 on advanced reasoning at roughly three to four times less inference compute—an estimate that source says is based on active parameters and generated tokens, not measured serving cost. Sources disagree on ranking: Latent Space cites observers placing Beam near GLM-5.2 and behind DeepSeek V4 and GLM 5.3, whereas The Decoder says it approaches Qwen3.8-Max on coding benchmarks with Kimi K3 still ahead on raw performance. Latent Space also reports US training from scratch on 23.8 trillion tokens; the reinforcement-learning run used about 10,500 NVIDIA GB300 GPUs for about four weeks and more than 100 million attempts or rollouts, and Help Net Security says scores were still rising when it stopped.