Reflection
Reflection opens Beam, a 501B coding model, under Apache 2.0
Reflection has published the weights for Beam, a 501 billion parameter open model, and says it matches larger rivals on coding with far less inference compute.
Reflection has published the weights for Beam, its first open-weight model, under an Apache 2.0 license that lets companies run it on their own hardware. The company says Beam competes with larger open models on coding while using a fraction of the compute.
Beam Is a 501 Billion Parameter Coding Model
Beam is a sparse mixture-of-experts model with 501 billion total parameters, of which 23 billion are active for each token. Reflection built it for coding, reasoning and agentic work, and it takes text only. The model card lists a 1 million token context window, 256 experts and a June 30, 2026 knowledge cutoff.
Reflection announced Beam on Oct. 5 and said weights, a technical report, a model card and developer artifacts would follow “later this month.” The Hugging Face repository is now ungated and carries BF16, FP8 and NVFP4 versions. The model card still reads “Technical Report (coming).”
Reflection Reports Mixed Results Against Larger Rivals
Beam scores 80.9 on SWE-bench Verified, 80.1 on Terminal Bench v2.1, 65.5 on SWE Bench Pro v1 and 44.4 on DeepSWE v1.1, according to a table in Reflection’s Oct. 5 post, updated Oct. 8. Every figure is self-reported.
GLM 5.2 scores 81.0 on Terminal Bench v2.1 and 62.1 on SWE Bench Pro v1. Qwen 3.8 Max scores 86.6 and 67.7 on those two tests. Kimi K3 reaches 88.3 on Terminal Bench v2.1 and DeepSeek V4.1 Flash reaches 90.6.
Reflection says Beam “advances the Western open-weight frontier” and is competitive with GLM 5.2 and approaching Qwen 3.8 Max on coding and agentic tasks. It concedes that models like Kimi K3 “remain ahead on raw capability.”
The Efficiency Claim Rests on Fewer Reasoning Tokens
Reflection says Beam scores comparably to GLM 5.2 on advanced reasoning benchmarks while using 3 to 4 times less inference compute. That is Reflection’s estimate, built from Artificial Analysis and DataCurve data. It approximates compute as 2 times active parameters times mean generated tokens and leaves out prompt prefill and serving overhead. The Hugging Face card puts the gap at about 3 times fewer reasoning tokens.
A reasoning-effort setting trades shorter answers for longer reasoning. Reflection pretrained the model on 23.8 trillion tokens, then ran reinforcement learning for four weeks on 10.5K Nvidia GB300 GPUs. That run generated more than 100 million rollouts across about 1 million training environments. Reflection believes it is one of the largest by any open lab.
Hosted Access Starts at $1 per Million Input Tokens
OpenRouter listed Beam on Oct. 16 at $1 per million input tokens and $3 per million output tokens, with cached input at $0.10, according to its model list on Oct. 18. Prices get cut. Reasoning is mandatory there, with efforts from low to max (default max). Reflection also offers an API with self-serve keys on its platform.
Our View
A reasoning model bills for the thinking it does before it answers. A model that reaches the same answer with about a third of the tokens costs less on every request.
Developers paying per output token see that gap on the invoice. A company running the weights on its own GPUs spends less GPU time per task. Apache 2.0 lets it run them without negotiating a license.
Those savings depend on independent testing confirming Reflection’s token counts. Every benchmark in the post is self-reported, and the technical report is not yet out.