The Pulse
Reflection Releases Its 501B Beam Model for Coding Agents
Reflection AI has released Beam, a 501-billion-parameter open-weight model aimed at coding, reasoning and agentic tasks. The Apache 2.0 release includes a one-million-token context window, several checkpoint formats and support across cloud
AI.info Team · Written with AI ·
On October 16, Reflection AI released the weights for Beam, its 501-billion-parameter model for coding, reasoning and agentic work, after previewing the model earlier in the month. The weights are available on Hugging Face under the Apache 2.0 license, the company says. “Today, we have released Beam’s model weights on Hugging Face under an Apache 2.0 license,” Reflection AI wrote in its release announcement.
501 billion parameters, 23 billion active
Beam uses a sparse mixture-of-experts architecture: it has 501 billion parameters in total, with 23 billion active for each token. Reflection describes it as a text-in, text-out model built for coding, reasoning and long-running agent tasks. Its advertised context window reaches one million tokens, and the release includes BF16, FP8 and NVFP4 checkpoints.
The launch also includes a Rust-based coding harness and day-one support in Transformers, vLLM and SGLang, according to Hugging Face’s technical launch guide. That combination is intended to let developers connect the model to tools and software-development workflows, rather than use it only for single-turn code generation. Reflection says users can download and host the weights themselves, fine-tune Beam, or access hosted versions through its API and partners.
Benchmark results carry a hardware caveat
In figures supplied by Reflection and published in Hugging Face’s launch guide, the FP8 checkpoint scores 80.2% on SWE-bench Verified, 78.9% on Terminal Bench 2.1 and 78.4% on SWE-bench Pro v2-Hard. Results vary by checkpoint: the same table reports 81.6% on SWE-bench Verified for BF16 and 80.7% for NVFP4. Those are company-provided benchmark results, not an independent comparison.
Running Beam locally calls for substantial memory despite its 23-billion active parameters. Hugging Face’s guide says the NVFP4 version reduces the model’s weight storage from about one terabyte in BF16 to roughly 280 gigabytes. It recommends four B200 GPUs for serving with room for long contexts and concurrent requests; smaller deployments face tighter memory limits. A one-million-token context therefore does not mean the model can use that context cheaply on ordinary workstation hardware.
More than a download
Reflection says more than 40 partners and integrations support the launch, spanning cloud platforms, inference providers and developer tools. Its list includes Amazon Web Services, Google Cloud, Microsoft Azure, Nvidia NIMs, Ollama, OpenCode and OpenRouter. Dell and HPE are listed for on-premises deployments, including sovereign and air-gapped environments.
The company frames the release as an open stack around the model, including deployment containers, inference integrations, customization pipelines and agent harnesses. That pitch matters because downloadable weights are only one part of operating a coding agent: teams also need hardware, serving software, evaluation and ways to connect a model to code repositories and tools. Beam’s Apache 2.0 license permits broad use of the weights, while the model’s practical reach will depend on whether developers can run and adapt it within their infrastructure.