Update: Reflection unveils Beam, a 501B Apache-2.0 open coder
Weights and FP8/NVFP4 quants drop later this month, scoring near GLM-5.2 at a claimed 3-4x less inference compute. Early access is open now.
Copy markdown
501B total, 23B active, Apache 2.0
Beam is a sparse mixture-of-experts pretrained on 23.8T tokens with a 1M-token context, shipping under a fully permissive Apache-2.0 license. Commercial use, self-hosting, and fine-tuning are all on the table - rare for a model this size.
77.2 SWE-Bench Pro v2-Hard, 80.1 Terminal-Bench
Reflection puts Beam near GLM-5.2 and Qwen 3.8-Max on coding and agentic work, while trailing DeepSeek V4.1 Flash and Kimi K3 on some tasks. Frontier-adjacent and open - not frontier-beating, but you can own the weights.
A claimed 3-4x cheaper to run than GLM-5.2
At comparable reasoning scores, Reflection claims 3-4x less inference compute, with FP8 and NVFP4 quantized builds on the way. That efficiency number is what decides whether self-hosting beats paying per token.
Weights land 'later this month' - waitlist is live
Only early access is open today; the weights, technical report, model card, and developer artifacts ship later in October under Apache 2.0. Sign up now if you want to test it the day it drops.