Reflection releases Beam, a 501B open model for coding and agents

A new frontier-scale open model, designed for long-context reasoning and agentic work, is now available for teams to own and customize. Beam, a 501-billion-parameter mixture-of-experts model from the new US lab Reflection, has been released on Hugging Face under the Apache 2.0 license, according to an announcement on the Hugging Face blog.

What is the Beam model and who is it for?

Beam is a text-only, decoder-based large language model built specifically for coding and agentic use cases, such as software development and autonomous task execution. According to the source, it is intended for teams that want a frontier-scale model to run on their own hardware or cloud infrastructure and wish to customize it for specific needs. The model is a sparse mixture-of-experts (MoE) architecture, which means that while it contains 501 billion total parameters, only 23 billion are active for any given token, allowing it to run at the speed of a much smaller model while retaining the capacity of a massive one.

What are the key technical specifications of Beam?

The model features a 1-million-token context window, a significant capability for processing long documents or maintaining extensive conversation histories. Its architecture includes 52 layers, with a hybrid attention system: most layers use a sliding-window attention over the previous 512 tokens, while every sixth layer and the final layer attend to the full context. This design, combined with a unique "scalable-softmax query scaling" technique in global layers, helps manage computational load for long sequences. The model also includes a custom residual stream that maintains a constant scale from input to output. It supports several precision formats, including BF16, FP8, and a 4-bit format (NVFP4) optimized for NVIDIA's latest Blackwell GPUs.

How does Beam perform and what is required to run it?

Benchmark results provided by the Reflection team show strong performance in agentic coding, reasoning, and tool-calling tasks. For example, on SWE Bench Pro v2-Hard, Beam scores between 76.5% and 78.6% across different precisions. On the AIME 2026 reasoning benchmark, it achieves scores around 97-98%. Running the model requires substantial hardware. The full BF16 checkpoint is about 1 terabyte. The more practical NVFP4 checkpoint reduces this to roughly 280 GB. The blog notes that a node with at least four NVIDIA B200 GPUs (720 GB of total memory) is recommended for serving the model with room for long contexts and concurrent requests, though it can also be run via serverless inference providers.

How can developers start using and building with Beam?

The model has day-zero support in popular inference libraries like Hugging Face Transformers, vLLM, and SGLang. The Hugging Face blog provides code snippets for running Beam using the text-generation pipeline or lower-level APIs with AutoModelForCausalLM. The model's tokenizer includes a built-in response template and parser that automatically structures outputs, separating reasoning, answers, and tool calls. Reflection and Hugging Face are also running a contest, offering $10,000 in Hub credits for the best projects built with Beam, such as demos on Spaces, quantized versions, synthetic datasets, or benchmarks.

What does Beam's release mean for AI creators and researchers?

The release of a 501B-parameter model under a permissive Apache 2.0 license represents a significant shift in accessible AI capability, particularly for specialized domains like coding and autonomous agents. It provides organizations with an alternative to closed, API-only frontier models, offering full ownership and the ability to fine-tune on private data. The model's architecture, especially its efficient long-context handling and built-in support for structured outputs like tool calls, sets a new standard for open models aimed at practical, complex applications beyond simple chat.

How can Grafy users explore models like Beam?

While Grafy does not currently host the 501B Beam model directly due to its size, users interested in the capabilities of large, open models for coding and research can explore similar frontiers within the platform. Grafy's multi-model Chat app allows for side-by-side comparison of many leading cloud models, providing a practical way to assess reasoning and coding abilities. For deeper technical analysis, the Reader app can be used to upload and interrogate the Beam model card, research papers, or benchmark reports, using AI to summarize findings and compare specifications. Developers can also prototype agentic workflows in Grafy Code or plan multi-step AI projects in the Graph.

Frequently asked questions

How is Beam different from other large open models like Llama?

Beam is specifically architected for long-context (1M tokens) reasoning and agentic work, with built-in support for tool calling and a unique hybrid attention system. Its 501B MoE design offers the capacity of a frontier model but activates only 23B parameters per token for efficiency.

Can I run Beam on consumer hardware?

No. According to the source, serving the model practically requires multiple high-end data center GPUs, such as four NVIDIA B200s. The smallest checkpoint (NVFP4) is ~280 GB. Users are advised to use serverless inference providers for testing.

What can I build with Beam's tool-calling feature?

The model supports a dedicated grammar for tool calling, allowing developers to build agents that can reliably interact with external APIs and functions. This is designed for creating automated coding assistants, research agents, or workflow automation systems.

---

Source: Hugging Face Blog — Welcome Beam: a 501B open model you can own, customize, and scale