Beam

Beam by Reflection AI is an AI developer tool in the open-weight large language model category, built for coding, reasoning, and agent workloads. It uses a sparse Mixture-of-Experts architecture with 501 billion total parameters but only 23 billion active per token, which the team cites as delivering 3-4x less inference compute than comparably capable models while holding competitive benchmark scores. The model supports up to a 1 million token context window and a configurable reasoning-effort parameter, so developers can trade response length against task difficulty. It is trained for tool use, demonstrating web search, code execution, and API calls, and reports results on software-engineering benchmarks such as SWE-Bench Pro and Terminal Bench alongside reasoning tests like AIME and GPQA. Training combined 23.8 trillion pretraining tokens with large-scale reinforcement learning across custom environments. Reflection plans to release the weights under the Apache 2.0 license with documentation and developer tooling; access is currently through an early-access signup. For teams weighing open-weight models to self-host for coding agents, Beam is a credible entry in the best-AI-for-agentic-coding conversation BestAIFor tracks. It targets enterprise developers who need efficient inference for software engineering and agentic systems.

Self-hosted coding, reasoning, and agentic workflows including software engineering and terminal tasks.