Beam is the first open-weight model from Reflection. On October 16, 2026, Reflection formally released Beam, and Ollama announced the same day in its official blog that it is a day-zero partner: the model has roughly 501B total parameters in a mixture-of-experts design, activates only about 23B parameters per token, offers a 1 million token context window, and ships its weights under the Apache 2.0 license.
Big parameters, small activation, aimed at long jobs
Beam's trade-off is explicit. A large total parameter count stores more knowledge and capability, while activating only a small slice per step keeps inference cost down. It is not aimed at single question-and-answer exchanges, but at work that runs for a long time: writing and maintaining code across a codebase, calling tools over many steps, and pushing a complex problem forward through logs and documents. Input and output are text, with code, tool calls, logs, documents and structured outputs at the center, and math and science reasoning also appear in the official list of target workloads. That contrasts with the Muse Spark Open weights covered here just before: two large US open-weight models, one leading with multimodality and local deployment, while Beam bets everything on coding and long-horizon agents.
The inference ecosystem showed up on day one
Whether an open-weight model runs on day one often decides whether the first wave of developers stays. Beam's supporting moves are unusually complete: Ollama Cloud offers it free for a limited time with full zero data retention, so developers can try it from a terminal without downloading hundreds of gigabytes of weights, and both vLLM and SGLang announced support the same day, with SGLang saying it already powers Reflection's own production API. Teams that want to self-host do not have to wait for community ports, and teams that only want to evaluate it get a near-zero-cost trial window.
Who should try it now, and who can wait
Teams building coding agents, automation or long tool-calling chains should put Beam on their evaluation list now: the scale, context length and license are all friendly to commercial use, and the only real unknown left is the pass rate on their own tasks. If your workload is image or video generation, or you only need a chatbot-grade model, Beam's text-only design will not help. Next to watch is community testing after the weights land: even with only 23B active, a 501B-class model still has a real deployment bar, and quantized releases plus measured throughput will decide whether it stays on leaderboards or actually enters workflows.
