Beam 1.0
Beam has 501.3 billion total and 23.4 billion active parameters. It uses fifty-two layers, with gated grouped-query attention and sliding-window attention.
Training and model formats
Beam was pretrained on 23.8 trillion tokens, followed by large-scale reinforcement learning. Reflection released the instruction-tuned model in several numeric formats. These formats do not represent independently trained model families.Reflection: Beam training announcementReflection: Beam weight repository
Reasoning and API behavior
The API supports Low, Medium, High, XHigh and Max effort. Reasoning cannot be disabled. Previous assistant messages must include their reasoning_content. Reasoning tokens count toward the output budget, so too small a budget can leave no final answer.Reflection: Reasoning and conversation requirements
Serving packages
The official collection contains BF16, FP8 and NVFP4 weights and a DFlash2 draft checkpoint. The draft supports speculative decoding with Beam rather than replacing the main model.Reflection: Beam 1.0 release collection
How to use it well
Practical guidance. Budget memory for the full checkpoint and long conversations. Verify tool execution and final outputs in the target application; fewer generated reasoning tokens do not by themselves establish better task performance.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable |
| Model identifiers | Beam-1.0-501B-A23Breflection/Beam-1.0-501B-A23B |
| Context window | 1,048,576 tokens API context includes prompt and generated output Reflection: API model catalog |
| Maximum output | 131,072 tokens Published model maximum; account and request limits can be lower Reflection: API model catalog |
| Knowledge cutoff | June 30, 2026Reflection: Beam model card |
| Modalities | Text input and output; developed mainly for EnglishReflection: Beam model card |
| Weight license | Apache 2.0Reflection: Beam weight repository |