Reflection's Beam becomes the most capable open-weight model built outside China

AI startup Reflection has announced Beam, its first open-weight model, built for coding, logical reasoning and agentic tasks. The mixture-of-experts model activates 23 billion of its 501 billion total parameters per token, a design intended to keep compute costs relatively low.

According to Reflection, Beam matches GLM 5.2 on demanding reasoning tasks while using three to four times less compute, and comes close to the much larger Qwen3.8-Max on coding and agent benchmarks. The company says stronger open models such as Kimi K3 still beat Beam on raw performance, and that it is already training a successor aimed at closing the remaining gap to top-performing open models. Reflection is targeting businesses that use AI for coding and automated workflows but need to control operating costs.

Beam combines large-scale pretraining with a compute-heavy reinforcement learning phase. Reflection says it ran 10,500 Nvidia GB300 GPUs for more than four weeks during that phase, describing it as one of the largest training runs any open lab has done. The company says performance kept improving through the end of the run without plateauing, even past 80 million rollouts. A tunable parameter lets users control how thoroughly the model reasons, trading compute cost against output quality.

Reflection reported what it calls emergent capabilities: during an RL mix of reasoning, software engineering and terminal tasks, Beam improved at web browsing although no browsing tasks were in that mix. With web access, the company says, the model independently learned to query other language models and pull documents from external services. Beam is text-only but can process content from other media if represented as text.

For safety and alignment, Reflection trained a second model and merged it with Beam, with guidelines ranging from hard rules to factual accuracy, admitting uncertainty and a direct, thorough response style. Safety test results and open-sourced evaluation methods are planned for a technical report.

A technical report, developer documentation and model weights under the Apache 2.0 license are set to ship later this month. Beam is still undergoing final safety testing, with an early version available to select users.

Reflection was founded in 2024 by former Google Deepmind researchers Misha Laskin and Ioannis Antonoglou. It launched in March 2025 with $130 million in seed funding, released the Asimov codebase agent in summer 2025, and raised $2 billion at an $8 billion valuation in October 2025, with Nvidia among the investors. The company publishes model weights but keeps training data and pipelines proprietary, and has signed billion-dollar compute deals with SpaceX and cloud provider Nebius.