Reflection has released Beam, its first model, on Hugging Face. The text-only model is built for coding, long-context reasoning and agentic tasks. It has 501 billion parameters, a context window of one million tokens and an Apache 2.0 license.
Beam uses a mixture-of-experts design. It activates about 23 billion parameters for each token while keeping the full model in memory. That makes memory capacity a practical consideration for teams planning to run it on their own hardware.
According to Hugging Face, Beam’s BF16 checkpoint is about 1 TB. An NVFP4 version reduces the weight storage to about 280 GB for NVIDIA Blackwell GPUs. FP8 checkpoints are also available for older GPUs, including Hopper and Ampere chips. Hugging Face recommends four B200 GPUs for NVFP4 serving to leave room for long contexts and concurrent requests.
Several platforms and software tools support Beam
Beam has support in transformers, vLLM and SGLang. It is also available through Hugging Face Inference Providers, powered by Baseten, Fireworks and Together. Teams can use supervised fine-tuning and reinforcement learning recipes to adapt the model.
Hugging Face reports that, after supervised fine-tuning on Multilingual-Thinking, Beam used the requested reasoning language on 16 of 16 held-out prompts. The base model did so on 3 of 16. The company says this small evaluation measures language adherence, not whether the reasoning was correct.
Reflection is offering $10,000 in Hub credits for the best projects built with Beam and shared on the Hub. Beam is available now on Hugging Face and through the listed Inference Providers.
