Reflection releases Beam, a 501B open model under Apache 2.0

Reflection, a new artificial intelligence lab based in the United States, released its first model, Beam, on Hugging Face on October 16, 2026. The mixture-of-experts model has 501 billion parameters in total, of which about 23 billion are active for each token. It is licensed under Apache 2.0, which permits commercial use.

Beam takes text input only and offers a context window of 1 million tokens. Reflection positions it for software engineering and agentic workloads, aimed at teams that want a frontier-scale model they can run and customize on their own hardware or cloud. The release includes a dedicated open-source harness written in Rust that the lab says gets the best performance from the model.

Architecture and hardware

The model has 52 layers. All but the first three use a mixture-of-experts block with 256 experts, eight of which are routed per token. Routing follows the group-limited scheme introduced by DeepSeek-V3, though Beam drops the shared expert. Forty-three layers use sliding-window attention over the previous 512 tokens, while nine attend to the full context, which keeps cache growth to roughly 36 KiB per token, about a sixth of what an all-global design would need.

Because every expert stays in memory, serving Beam requires the memory footprint of a 501B model while running at speeds closer to a 23B one. The BF16 checkpoint is roughly 1 TB. Reflection also provides FP8 weights suited to Hopper and Ampere chips and an NVFP4 version that stores most parameters in 4-bit precision for NVIDIA Blackwell GPUs. Support in transformers, vLLM and SGLang was available from launch, along with recipes for supervised fine-tuning and reinforcement learning.

Benchmarks

According to figures supplied by Reflection, the BF16 version scores 81.6% on SWEBench Verified, 80.1% on Terminal Bench v2.1 and 66.1% on SWE Bench Pro v1. The company's numbers show only small differences between full and reduced-precision checkpoints. These results come from the developer and have not yet been independently verified.

Reflection is also offering $10,000 in Hugging Face Hub credits to the best community projects built on Beam, from demos to quantizations and benchmark studies. The launch adds another large open-weight option for organizations seeking a coding-focused model they can deploy on infrastructure they control.