Reflection launches Beam, a 501B open AI model
Reflection has released Beam, a 501B-parameter mixture-of-experts model designed for coding, long-context reasoning and agentic workloads. The Apache 2.0-licensed model is intended for teams that want to run and customize a frontier-scale system on their own infrastructure.
Radwan AbdulhadiEditor and technology writer
Open release targets private deployment
Beam is the first model from Reflection, a new US-based lab, and is being released as an open model through Hugging Face. The company describes it as a 501B-parameter mixture-of-experts large language model with 23B parameters active per token. It accepts text input, produces text output and has a 1M-token context window. The model is released under the Apache 2.0 license, allowing teams to own, customize and deploy it on their own hardware, infrastructure or cloud.
Reflection is positioning Beam for coding and agentic use cases rather than as a general-purpose consumer product. The release emphasizes control over deployment and customization, with the company presenting the model as an option for teams that want frontier-scale capabilities without relying exclusively on a hosted service.
Sparse architecture supports long contexts
Beam uses a sparse mixture-of-experts architecture in which only a fraction of the network processes each token. Its decoder has 52 layers and 256 experts in most of its mixture-of-experts layers, while eight routed experts are selected for each token. This gives the model the computational profile of a much smaller model while retaining the capacity of a 501B-parameter system. All experts must remain in memory, however, so deployment still requires memory for the full model.
The architecture combines sliding-window and global attention to limit the cost of long sequences. Forty-three layers attend to a fixed window of the previous 512 tokens, while nine global layers can attend to the full context. Hugging Face says this design makes the key-value cache grow by about 36 KiB per token in BF16 with the Transformers dynamic cache, compared with 208 KiB per token if every layer retained the full sequence.
The model also uses grouped-query attention, scalable-softmax query scaling in global layers and a residual design intended to keep the stream at a constant scale through the network. The first three layers use dense SwiGLU blocks; the remaining 49 use expert blocks. Beam applies selected computations, including routing, normalization and the final logit softcap, in FP32.
Reasoning and agent features
Beam supports reasoning that can remain in context across turns through a preserve_thinking capability. Users can select among five reasoning levels—low, medium, high, xhigh and max—with medium set as the default. The model also includes tool calling through a dedicated grammar, features aimed at workflows in which the model must plan, invoke tools and continue work across multiple interactions.
The release includes a speculative decoding drafter model based on DFlash2. Reflection and Hugging Face have also prepared recipes for inference, supervised fine-tuning and reinforcement learning, including full-parameter reinforcement learning with Miles. These materials are intended to help teams adapt Beam to their own applications rather than only run the base checkpoint.
Hardware formats and serving options
Beam is available in several precision formats. The BF16 checkpoint is about 1 TB in safetensors format, while the NVFP4 version stores most parameters at 4-bit precision and reduces the model to roughly 280 GB. NVFP4 is intended for NVIDIA Blackwell GPUs, while FP8 checkpoints are positioned for Hopper and Ampere systems. The available checkpoints include [BF16](https://huggingface.co/reflection/Beam-1.0-501B-A23B), FP8 and NVFP4 variants.
Hugging Face says two B200 GPUs can hold the NVFP4 weights, although four B200s leave more room for activations, the key-value cache, long contexts and concurrent requests. The model can also be served on an eight-B200 node, two B300 GPUs or eight RTX Pro 6000 GPUs, according to the release. Teams without suitable Blackwell hardware can use serverless Inference Providers to test the model.
Beam has day-one support in Transformers, SGLang and vLLM. Reflection also supplies a dedicated open-source Rust-based custom harness intended to deliver the model's best performance. The harness is documented separately by Reflection, alongside the deployment recipes for the downstream serving libraries.
Credit program encourages community projects
Reflection is offering $10,000 in Hub credits to projects that use Beam. Eligible work can include a Spaces demo, a quantized version for specific hardware, a synthetic dataset, an reinforcement-learning environment or a benchmark. Participants must create a model, dataset or Space repository and share it on the Beam model page; Reflection says it will collect the links, select the best projects and send credits to their Hub accounts.
The program gives developers a direct way to test the model's performance and adapt it to specialized workloads. It also extends the release beyond a checkpoint by encouraging shared tooling, evaluations and applications around Beam's long-context, coding and agentic capabilities.
Sources
- Hugging Face BlogWelcome Beam: a 501B open model you can own, customize, and scale
Continue chronologically
Related stories
Hugging Face Transformers Now Supports llama.cpp GGUF Models
Hugging Face has added native support for GGUF model files to its Transformers library, enabling developers to run quantized AI locally.
Mistral Large 4 Released on Hugging Face
Mistral has launched Mistral Large 4, an open-weight Mixture-of-Experts multimodal model featuring 1T total parameters and strong cybersecurity capabilities.
IBM Announces Granite Speech 5.0 Collection
IBM unveils Granite Speech 5.0, introducing a new speech-language model, a faster multilingual Turbo CTC, and an updated 2B Plus model for enterprise audio.
AWS Introduces Support for Reflection AI's Beam
Amazon Web Services announces support for Reflection AI's new open-weight agentic coding model, Beam, on Amazon SageMaker.
