Reflection releases Beam, a 501-billion-parameter open-weight model aimed at coding agents

Reflection AI has released Beam, a large open-weight model designed for coding, reasoning and agentic tasks. The company says developers can download the weights under the Apache 2.0 license, self-host the model or use hosted offerings from Reflection and its partners.

Beam is a sparse mixture-of-experts model with 501 billion parameters in total and 23 billion active for each token. That design limits the model’s per-token computation, but it does not remove the need to hold its full set of weights in memory. Hugging Face’s technical overview says the BF16 checkpoint is about 1 terabyte; its NVFP4 version is about 280 gigabytes, according to the same post.

Reflection is positioning the release as more than a set of weights. It says the launch includes more than 40 partners and integrations covering inference, deployment, evaluation, customization and agent runtimes. The company also released Mirror CLI, an open-source coding and agent harness that lets developers inspect and approve tool calls, steer work and run parallel agents in separate worktrees.

The weights are available in BF16, FP8 and NVFP4 formats, with day-one support listed for Transformers, SGLang and vLLM. Reflection says it has also released a speculative-decoding drafter model and worked with partners on serving configurations and fused kernels. These are vendor descriptions of the release; the supplied materials do not include independent performance testing of those optimizations.

The model’s size makes infrastructure a central part of the announcement. AWS says Beam requires a multi-GPU deployment and is available through SageMaker JumpStart in the US East (N. Virginia) and US East (Ohio) regions. Its deployment guidance recommends sizing for weight memory, the KV cache and runtime overhead, then validating the setup with load tests.

Reflection says Beam is available through a broad group of cloud, inference and software partners, including Amazon Web Services, Google Cloud, Microsoft Foundry, CoreWeave, Fireworks AI and Together AI. It also lists Dell and HPE as options for on-premises, sovereign and air-gapped deployments. Availability and configurations may vary by provider; AWS, for example, identifies two launch regions for its JumpStart offering.

The release materials include benchmark scores supplied by Reflection, spanning coding, reasoning and tool-use tasks. Hugging Face reproduces those figures but attributes them to Reflection, so they should be read as company-reported results rather than independent confirmation. The source materials do not establish how Beam compares with other models under a common, independently run evaluation.

Reflection says it used roughly 1.3 billion sandboxes across reinforcement-learning training, evaluation and ablation work on Beam. The company also describes early developer projects, including an escape-room game built with Beam and a comparison of small coding projects. Those examples show what early users have tried, but they are not a substitute for broader evidence about reliability on production coding workloads.