# Reflection releases open-weight Beam 1.0 with 501B parameters and 1M-token context

Digest AI · Generative AI & Models · published 2026-10-16T14:22:08Z · updated 2026-10-16T15:23:10Z

Canonical: https://digestai.news/story/reflection-releases-beams-open-weights-under-apache-2-0-license

## Summary

Reflection announced Beam 1.0, an open-weight language model with 501 billion parameters and 23 billion active parameters using a sparse mixture-of-experts architecture. The model supports a 1 million-token context window and is optimized for English, with instruction-tuned versions in BF16, FP8, and NVFP4 formats. Beam claims token-efficient reasoning, outperforming leading open models with three times fewer inference tokens, and excels in software engineering, agentic tool use, and advanced reasoning tasks like math and science problem-solving.

The model was trained on diverse datasets, including web content, source code, technical documentation, and proprietary vendor data. Reflection emphasizes its safety evaluations, including Open Safety Framework tests and its in-house Iris Safety Evaluations (ISE). Beam is licensed under Apache 2.0 and intended for commercial and research use, with explicit permissions for fine-tuning, distillation, and security research. The model has already been downloaded 440 times on Hugging Face since its release.

## Key points

- Beam 1.0 has 501B total parameters and 23B active parameters with sparse MoE architecture
- Supports 1M-token context window and optimized for English with instruction-tuned versions in BF16/FP8/NVFP4
- Achieves 81.6% on SWE-bench Verified, 90.6% on Diamond GPQA, and 80.1% on Terminalbench 2.1

## Why it matters

Beam’s efficiency in token usage and strong performance in coding, reasoning, and agentic tasks could lower barriers for researchers and developers working with large models. Its open-weight release and permissive license may accelerate innovation in AI agents and automation.

## Sources

1. [reflection/Beam-1.0-501B-A23B · Hugging Face](https://huggingface.co/reflection/Beam-1.0-501B-A23B) (huggingface.co, 2026-10-16, primary source)
2. [Introducing Reflection Beam: an open-weight model for agentic coding on AWS](https://aws.amazon.com/blogs/machine-learning/introducing-reflection-beam-an-open-weight-model-for-agentic-coding-on-aws) (AWS Machine Learning Blog, 2026-10-16, primary source)
3. [Reflection Releases Beam’s Open Weights Under Apache 2.0 License](https://unite.ai/reflection-releases-beams-open-weights-under-apache-2-0-license) (Unite.AI, 2026-10-16)
4. [Beam 1.0 - a reflection Collection](https://huggingface.co/collections/reflection/beam-10) (huggingface.co, 2026-10-16, primary source)

Part of the developing story: [Reflection AI's Beam Model Saga](https://digestai.news/thread/reflection-unveils-beam-an-open-model-it-says-rivals-glm-5-2-on-reasoning-while) (2 stories)

## Cite

Digest AI, "Reflection releases open-weight Beam 1.0 with 501B parameters and 1M-token context", 16 October 2026, https://digestai.news/story/reflection-releases-beams-open-weights-under-apache-2-0-license

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/reflection-releases-beams-open-weights-under-apache-2-0-license.json