Reflection, a new US-based AI lab, has released Beam, a 501 billion parameter mixture-of-experts (MoE) language model with 23 billion active parameters per token. The model is fully open-source under Apache 2.0 license, offering frontier-scale capabilities for teams wanting to deploy on their own infrastructure.
Technical Architecture
Beam uses a sparse MoE architecture where each token is computed by a small fraction of the network, enabling it to run at the speed of a much smaller model while maintaining the capacity of a 501B parameter model. The architecture includes:
- 52 layers with a hidden size of 5,120
- 256 experts per MoE layer with 8 routed per token
- 1 million token context window
- Sliding-window attention (512 tokens) in most layers, with full-context global attention every sixth layer
- Grouped-query attention with 48 query heads and 8 key-value heads
The model supports five reasoning effort levels: “low”, “medium” (default), “high”, “xhigh”, and “max”, allowing users to trade off speed for reasoning quality.
Checkpoint Formats and Deployment
Beam is available in three checkpoint formats:
- BF16: Full precision at approximately 1 TB
- FP8: Reduced precision for efficiency
- NVFP4: 4-bit quantization for NVIDIA Blackwell GPUs
The BF16 checkpoint requires significant memory (about 1 TB), but the model runs at the speed of a 23B parameter model since only 8 experts activate per token.
Ecosystem and Tools
Reflection has partnered with Hugging Face to provide day-0 support across popular libraries:
- Transformers: Native implementation available from launch
- SGLang and vLLM: Optimized inference support
- Miles: Full-parameter reinforcement learning framework
- Custom Rust harness: Purpose-built for optimal performance
The team is offering $10,000 in Hugging Face Hub credits for the best projects that use Beam, encouraging demos, quantizations, synthetic datasets, RL environments, and benchmarks.
Focus on Coding and Agents
Beam is specifically designed for coding and agentic use cases. It includes features like reasoning that preserves context across conversation turns and tool calling with a dedicated grammar. The model is positioned for teams wanting frontier-scale capabilities on their own hardware without relying on API calls.
The Apache 2.0 license means enterprises can fully own, customize, and scale the model for their specific needs—a significant departure from closed API-only approaches.