Reflection’s Beam pairs open weights with a substantial deployment challenge

Reflection AI has released Beam, an open-weight model aimed at coding, reasoning and agentic tasks. The company offers the weights under the Apache 2.0 license and says developers can self-host and customize the model or use it through Reflection and partner services. Reflection describes the launch as an open agent stack, not just a model release: it lists more than 40 partners and integrations across inference, deployment, evaluation, customization and agent runtimes.

Beam is a sparse mixture-of-experts model with 501 billion parameters in total and 23 billion active per token. That limits the parameters used for each token, but the full set of weights still has to be available to serve the model. Hugging Face’s technical overview puts the BF16 checkpoint at about 1 terabyte and the NVFP4 checkpoint at roughly 280 gigabytes. The overview also describes a one-million-token context window and a hybrid attention design intended to reduce per-sequence KV-cache growth compared with a model in which every layer attends to the full context.

Reflection lists BF16, FP8 and NVFP4 versions, a speculative-decoding drafter model, and day-one support for Transformers, SGLang and vLLM. It has also released Mirror CLI, an open-source coding and agent harness. Reflection says Mirror lets developers inspect and approve tool calls, steer an agent, resume conversations and run parallel agents in separate worktrees; the company says it trained Beam with Mirror while also testing it with other harnesses. These are product and compatibility claims from the company, not independent evidence that the model performs equally well across those setups.

The model’s scale makes deployment a significant part of the story. AWS says Beam requires a multi-GPU deployment and identifies SageMaker JumpStart availability at launch in US East (N. Virginia) and US East (Ohio). AWS offers a managed real-time endpoint path as well as a more configurable HyperPod route using JumpStart, SGLang or vLLM. Its guidance stresses planning for checkpoint storage, weight memory, KV cache and runtime overhead, and testing a deployment against the workload it is meant to serve.

Reflection lists hosted options across cloud and inference providers, alongside Dell and HPE deployments for on-premises, sovereign and air-gapped environments. It also says the launch ecosystem includes deployment containers, inference integrations and customization pipelines. Availability is provider-specific: AWS’s stated two-region JumpStart launch, for example, is not the same as general availability across all providers or environments.

Reflection’s reported benchmark results cover coding, reasoning and tool-use tasks. In the figures reproduced by Hugging Face, Beam’s BF16 version scores 81.6% on SWE-bench Verified, 80.1% on Terminal Bench 2.1 and 51.1% on Humanity’s Last Exam with tools. Hugging Face attributes the results to Reflection; the supplied materials do not provide an independent, common evaluation establishing how Beam compares with other models.

Reflection says its largest reinforcement-learning runs involved roughly 1.3 billion sandboxes across training, evaluation and ablations, and says it is releasing recipes and integrations for developers who want to post-train Beam. The company has also highlighted early projects built with the model, including a tool-using escape-room game. Those examples illustrate possible uses, but do not by themselves establish reliability on production coding or other long-running agent tasks.