Reflection AI announced on Friday, October 16, 2026, the availability of the Beam 1.0 weights, its first open-weight model. The release is under the Apache 2.0 license on Hugging Face and includes BF16, FP8, and NVFP4 variants, plus a drafter model for speculative decoding.
Beam is a sparse Mixture-of-Experts with 501 billion total parameters and about 23 billion active per token. It was designed for advanced coding, reasoning, and long-horizon agentic workloads, with an effective context of up to 1 million tokens. The company, founded by former Google DeepMind researchers, had introduced the model earlier this month; now the weights, model card, and safety report are public.
Open ecosystem
In addition to the weights, Reflection is launching what it calls an “open operating system for agents.” The model acts as the kernel, while upper layers — agentic integrations, containers, customization pipelines, and inference tools — are open. More than 40 partners offer day-zero support, including AWS, Google Cloud, Microsoft Azure, NVIDIA NIMs, Hugging Face, vLLM, SGLang, Ollama, and LM Studio.
The Mirror CLI, the open-source harness the team used internally to train the model, has also been released. It lets developers run Beam from the terminal, inspect actions, approve tool calls, and adapt the workflow with skills and MCP.
According to the official announcement, developers can download the weights, self-host, evaluate, customize with fine-tuning or RL, and connect to existing agentic systems. The company highlights efficiency: the MoE design enables competitive performance with larger models while using less inference compute on reasoning tasks.
The launch reinforces the competition for high-capability open models, especially in coding and agents, at a time when American labs are seeking alternatives to large Chinese models.
Sources
Transparency: This content was created, edited, or reviewed with the aid of artificial intelligence. The information was cross-checked with public posts on X and sources available on the internet. Consult the original sources for the full context.
By GeekikiBot
