Reflection AI: the Beam family of open models
Try open-weight models from U.S. lab Reflection AI, optimized for coding and agentic tasks, and served on Crusoe infrastructure.
Built for long-running agents
Beam is trained for coding, reasoning, and multi-step tool use. Crusoe’s inference engine, powered by MemoryAlloy™ technology, shares the KV cache across nodes, so agent loops that resend the same long context skip redundant prefill and return responses faster.
An open model built in the U.S., kept on Crusoe Cloud
Beam runs on Crusoe-operated infrastructure, so your prompts and outputs stay with Crusoe Cloud and are never sent to Reflection AI.
Serve without managing clusters
Deploy Beam models for inference using Self-Serve Deployments. Get a dedicated endpoint optimized for throughput or responsiveness based on your unique needs.
Fine-tune and deploy in Crusoe Intelligence Foundry
Frequently
asked questions
Reflection AI is a New York-based lab founded in 2024 by Misha Laskin and Ioannis Antonoglou, former Google DeepMind researchers. Backed by NVIDIA, Reflection AI set out to build a U.S. open frontier lab and released Beam, its first open-weight model family, in October 2026 under the Apache 2.0 license. On Crusoe Cloud, Beam runs on Crusoe-operated infrastructure with no customer data sent to Reflection AI. Request a Tailored Deployment for bespoke optimization and a dedicated endpoint.
Beam is open-weight under the Apache 2.0 license, which permits commercial use, modification, and redistribution without royalties. Reflection publishes the weights and research behind its models but keeps its training datasets and full training pipelines proprietary. For most teams, that means full freedom to deploy and fine-tune Beam, even without the ability to reproduce its training from scratch.
Beam is built for coding, reasoning, and agentic workloads, including code generation, debugging, multi-step problem solving, and agents that call tools across many steps. Current Beam models are text-only, so they suit software engineering, research, and automation tasks rather than image or audio work. Check each model’s details for current capabilities.
Beam uses a sparse mixture-of-experts architecture with hundreds of billions of total parameters, of which only a small fraction activate per token. That design gives Beam large total capacity while keeping inference compute closer to a much smaller model, which lowers latency and serving cost for interactive and agentic workloads.
Experiment with Reflection AI models in Crusoe Intelligence Foundry. Deploy for high-performance inference optimized for throughput or responsiveness with Self-Serve Deployments, or request a dedicated, SLA-backed endpoint through a Tailored Deployment. Each model’s details show the services it currently supports.
