Reflection releases weights and agent tools for its 501B-parameter Beam model
Beam’s Apache 2.0-licensed BF16, FP8, and NVFP4 weights are now on Hugging Face, connecting the model to coding and long-horizon agent work through Mirror CLI and widely used inference frameworks.
AT A GLANCE
Key points
- The model activates 23 billion of its 501 billion total parameters per token.
- BF16, FP8, and NVFP4 weights were released under Apache 2.0.
- A one-million-token context window and Mirror CLI target long-horizon agent workflows.
What is changing?
Reflection released Beam’s model weights on Hugging Face under the Apache 2.0 license on October 16. The sparse mixture-of-experts model has 501 billion total parameters and activates 23 billion per token. Hugging Face’s model introduction lists a one-million-token context window for Beam, which uses text input and output.
BF16, FP8, and NVFP4 checkpoints are available alongside a drafter model for speculative decoding. Transformers, vLLM, and SGLang have day-one support. Reflection’s terminal-based Mirror tool lets developers inspect agent activity, approve tool calls, resume sessions, and manage parallel agents in separate worktrees.
The hardware threshold remains high: Hugging Face lists the BF16 weights at about 1 TB and the NVFP4 version at about 280 GB. Hosted providers make evaluation easier, but on-premises deployment still requires data-center-class resources, task-specific evaluation, and strict tool permissions. Reflection also published a model card and a safety technical report.
Explore the original sources
Source date: 16 October 2026.
Published by AI Radar: 16 Oct 2026, 16:12 (Istanbul, UTC+3).
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…