Reflection AI releases Beam weights under Apache 2.0

Reflection AI released Beam’s model weights on Hugging Face under the Apache 2.0 license. Developers can now download and self-host the files instead of relying only on hosted deployments.

Beam’s instruction-tuned model comes in BF16, FP8 and NVFP4 formats. Reflection says developers can evaluate it, adapt it to their own domains and connect it to agentic systems, while the company’s developer documentation provides separate paths for deployment and customization.

According to the Beam model card, Beam is a sparse mixture-of-experts text model with 501.3 billion total parameters, 23.4 billion active parameters, 52 layers and 256 experts. That architecture uses only part of the model for each token rather than activating every parameter. Reflection lists a one-million-token context window.

Reflection says Beam has launch-day support for Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and NVIDIA Dynamo. The company also released serving configurations and fused kernels intended, it says, to reduce latency and increase throughput.

The weights follow Reflection’s October 5 introduction of Beam, when the company said final red-teaming and evaluations were still underway and promised weights, a model card, a technical report and developer artifacts later in the month. The October 16 download release is distinct from both that introduction and Beam’s earlier availability through AWS SageMaker JumpStart.

For the BF16 version, Reflection reports company-published scores of 81.6% on SWE-bench Verified, 80.1% on Terminal Bench 2.1 and 97.7% on AIME 2026. Those results have not been independently reproduced under identical settings, and the available evidence does not establish how Beam performs in production deployments.

Reflection also claims Beam can reach comparable solution quality to leading open models while using about three times fewer inference-time reasoning tokens. The company says its largest reinforcement-learning run used roughly 1.3 billion sandboxes and peaked at 883,006 virtual CPUs across 23 clusters on Google Cloud and AWS. Neither the efficiency claim nor the training-scale figure has been independently verified.

Reflection’s public GitHub organization lists Apache 2.0 repositories for its Mirror coding harness and a Beam endpoint verifier, both updated on October 16. The company said self-distillation recipes would follow in the coming days, but they were not available at the reporting cutoff.