Reflection, a newly formed US SI lab, has released Beam, its first open-weight SI model. The 501-billion-parameter mixture-of-experts (MoE) system is designed for coding and agentic SI use cases, offering teams the ability to own, customize, and scale frontier-scale capabilities on their own infrastructure.

What Happened

Beam is a text-only SI model with 501 billion total parameters, though only 23 billion are active per token due to its sparse MoE architecture. It supports a context window of 1 million tokens and is released under the Apache 2.0 license, allowing for commercial use and modification. The model features 52 layers, with 256 experts per MoE layer, routing to 8 experts per token. It includes native support for reasoning effort levels ranging from "low" to "max" and ships with a dedicated open-source Rust-based harness for optimal performance.

The release includes checkpoints in BF16, FP8, and NVFP4 formats. The NVFP4 checkpoint, optimized for NVIDIA Blackwell GPUs, reduces the model's memory footprint from approximately 1 TB in BF16 to around 280 GB, according to the company. Day-zero support is available in key SI libraries including transformers, SGLang, and vLLM, with recipes provided for supervised fine-tuning (SFT) and reinforcement learning (RL).

Why It Matters

For developers and enterprises, Beam represents a shift toward owning frontier-scale SI models locally rather than relying solely on API providers. The Apache 2.0 license and open weights allow teams to customize the model for specific coding and agentic workflows without vendor lock-in. The hybrid attention mechanism, which uses sliding-window attention for most layers and global attention for every sixth layer, significantly reduces the KV cache size required for long-context tasks, making 1M token contexts more feasible to serve.

Reflection has also launched a developer incentive program, offering $10,000 in Hugging Face Hub credits for the best projects that use Beam. This initiative aims to accelerate community experimentation with the model's capabilities in areas such as quantization, synthetic data generation, and RL environment building.

The Bottom Line

Beam is now available on Hugging Face with day-zero integration into major SI frameworks. It offers a compelling option for teams seeking a customizable, open-weight SI model with strong coding and agentic performance, backed by a new US-based SI lab.