In brief
- Beam is presented as a mixture-of-experts (MoE) language model with 23 billion parameters active per token.
- The model accepts and produces text, with a context window of up to 1 million tokens.
- Released under the Apache 2.0 license, Beam is available with support for Transformers, SGLang and vLLM.
- According to its creator, the NVFP4 version requires around 280 GB of weight memory; this format runs natively on Blackwell GPUs.
Reflection’s 501-billion-parameter AI model, Beam, was released via Hugging Face on 16 October 2026. Available under the Apache 2.0 license, the text model is focused on long-context reasoning, coding and agent-based use cases. Beam accepts text input and generates text output only.
A large model with sparse computation
Beam uses a mixture-of-experts (MoE) architecture. While the model has 501 billion parameters in total, around 23 billion are active when processing each token. According to technical details shared by Hugging Face, the network has 52 layers; each MoE layer contains 256 experts, of which 8 are active per token. This approach uses only part of the network during computation, but all experts are kept in memory. As a result, the number of active parameters alone does not indicate the model’s total memory requirements.
Beam’s context window extends to 1 million tokens. Its architecture combines layers that operate on 512-token windows with layers that attend to the full context. According to Hugging Face’s estimate for BF16, the cache for a 1-million-token sequence would be 208 GiB if every layer stored every token; in Beam’s configuration, the figure is estimated at 36.08 GiB. These are technical estimates for a single sequence using a dynamic cache.
License, formats and software support
The model was released under the Apache 2.0 license. Hugging Face’s announcement says support for Transformers, SGLang and vLLM was available on release day. Beam’s response template supports handling reasoning content, answers and tool calls in separate fields. The model also offers five reasoning levels, with “medium” as the default.
Evaluations shared by the Reflection team compare BF16, FP8 and NVFP4 across various coding, reasoning and tool-use tests. Since these are the creator’s own results, the same performance should not be assumed across different hardware and workloads. Testing with relevant use cases is recommended before deployment.
How demanding are the hardware requirements?
BF16 weights take up around 1 TB in safetensors format. According to Hugging Face, the NVFP4 format reduces this to around 280 GB and runs natively on Blackwell GPUs; GPUs released before Blackwell do not support NVFP4. FP8 checkpoints are recommended for those GPUs.
Hugging Face notes that two B200 GPUs can fit the weights, but leave little room for activations and the KV cache; it recommends planning for four B200s for long contexts and concurrent requests. It also provides an example of running the FP8 version on eight H200s on Hopper GPUs. For those without Blackwell hardware, serverless trials are available through Inference Providers.
Architecture and visualization workflows
Beam’s text-based design and coding focus suggest that architecture firms and visualization teams may want to evaluate it primarily for text and coding tasks. For example, teams looking to develop custom scripts or tool-using workflows can test the model. However, the source does not mention a ready-made integration with architecture software or an image-generation feature.
When planning to run the model on your own infrastructure, consider GPU memory, the numerical format you intend to use and the cache requirements of long contexts together. Offices should also verify that their inference library supports Beam and is compatible with their existing hardware. Serverless trials are an option when local hardware is unavailable, but the source does not explain the pricing for this option.
Sources
1 sourceSource texts are not republished; short quotes are marked, everything else is our own summary and commentary.
Beam’s open license and support for Transformers, SGLang and vLLM offer useful flexibility for architecture firms developing their own tools. Its coding focus may appeal to teams looking to build custom scripts or text-based automations for visualization workflows; however, the model is not presented as a ready-made architectural design or image-generation solution.
For teams in Turkey, hardware and operating costs should be key considerations. An open license does not remove the high memory requirement; those without suitable GPUs can consider a serverless trial. Since the source does not specify the cost of this option, teams should assess the license terms, access costs and performance on real workloads before using it in production.
Frequently asked questions
Under which license was Beam released?
Beam was released on Hugging Face under the Apache 2.0 license.
How large is Beam’s context window?
The model supports a context window of up to 1 million tokens.
Which GPUs can run Beam?
According to the source, the NVFP4 format runs natively on Blackwell GPUs; FP8 versions are recommended for GPUs released before Blackwell. For Hopper GPUs, Hugging Face gives an example of running the FP8 version on eight H200s.
