Reflection AI has released the weights for the Beam model, its first open-weight model with 501 billion parameters. Backed by NVIDIA and founded by former Google DeepMind researchers, the company has published the weights on Hugging Face under the Apache 2.0 license, alongside a model card and safety report.
Beam is a sparse Mixture-of-Experts with 501 billion total parameters and approximately 23 billion active parameters per token. It was built specifically for coding, reasoning, and long-horizon agentic tasks, featuring a 1-million token context window and a knowledge cutoff of June 2026.
Why Beam is Turning Heads
The company positions the Beam model as a Western alternative to Chinese open-weights (such as those from DeepSeek, Moonshot, and Zhipu) that many developers already rely on for cost efficiency. With fewer active parameters per token, Beam aims to complete tasks using fewer tokens, effectively reducing the cost per result. The weights are released in BF16, FP8, and NVFP4, with native support in Transformers, vLLM, SGLang, Unsloth, and others.
In addition to the Beam weights, Reflection is launching an ecosystem featuring over 40 partners (AWS, Google Cloud, NVIDIA, CoreWeave, Hugging Face, etc.) and tools like the Mirror CLI for agentic development. Developers can self-host, fine-tune with RL, and use hosted APIs.
Context and Impact
The release of Beam arrives amid fierce competition in open models for agents and coding. While Chinese labs lead in efficiency and the sheer volume of open-weights, Reflection is betting on a high-performing Western model with a permissive license and a focus on token efficiency. Published benchmarks show strong results in SWE-Bench and other coding and agent tasks.
It is still early to gauge large-scale adoption of Beam, but making the weights available allows the community to test, customize, and compare them directly against existing alternatives.
Sources
Transparency: This content was created, edited, or reviewed with the assistance of artificial intelligence. Information was cross-referenced with public posts on X and available internet sources. Consult the original sources to verify the full context.
By GeekikiBot
