Reflection AI released the open weights of Beam, its first large language model, on Thursday, October 16, 2026. The model is a Mixture-of-Experts with 501 billion total parameters and about 23 billion active per token, focused on coding, reasoning, and long-horizon agentic tasks.
The weights are available on Hugging Face under the Apache 2.0 license, in BF16, FP8, and NVFP4 formats, accompanied by a speculative decoding drafter. The company also published the model card with final benchmarks and the Beam Safety Technical Report.
What changes
Previously, Beam was in early access via waitlist after the October 5 announcement. Now any developer can download, self-host, evaluate, and customize the model. Reflection launched an ecosystem with over 40 partners, including AWS, Google Cloud, NVIDIA NIMs, Ollama, OpenRouter, and others, and opened a public beta of its own API.
The Mirror CLI, the open-source harness on which Beam was trained, was also released, along with integrations with OpenCode, Hermes, and others. Fine-tuning and RL recipes are available with partners such as Hugging Face, NVIDIA, and Axolotl.
Why it matters
Beam positions itself as a Western open-weights alternative, competing in inference efficiency with larger models like GLM-5.2, using about 3 to 4 times less compute on some reasoning benchmarks, according to the company. In benchmarks reported by Reflection, it scores 80.9 on SWE-bench Verified and 80.1 on Terminal Bench v2.1, trailing Chinese leaders like DeepSeek V4.1 Flash and Kimi K3 on harder agentic tasks.
NVIDIA is already an investor in Reflection and was in talks to deepen the investment or acquire the company, as previously reported. The weights release strengthens the open ecosystem and GPU demand.
Sources
Transparency: this content was created, edited or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet.
By GeekikiBot