Reflection AI's Beam Model Achieves High Efficiency
Reflection AI trained its 501-billion-parameter Beam model using 10,500 NVIDIA GB300 GPUs, processing 46.4 million sandboxes daily over four weeks.
Reflection AI has unveiled Beam, a 501-billion-parameter open-weight AI model, claiming significant efficiency gains over rivals. The U.S. Startup, backed by Nvidia, trained Beam over four weeks using a massive deployment of 10,500 NVIDIA GB300 GPUs, processing 46.4 million sandboxes per day.
Beam employs a sparse Mixture of Experts (MoE) architecture where only 23 billion parameters are activated for any given computation. This design is central to its efficiency. The model incorporates several architectural innovations to counteract policy staleness and improve training. These include SandwichNorm, attention gating, depth-based residual scaling, uniform expert utilization, a controlled residual stream, and FP32 residual accumulation.
Reflection states Beam delivers performance on par with major open-weight models from both the United States and China. For advanced reasoning benchmarks, it showed performance comparable to China's GLM-5.2. The company claims Beam handles coding and AI agent tasks at a lower expense than competing models.
Specifically, Reflection AI said it reduced the compute required for reasoning to one-third to one-quarter that of competing models while delivering higher performance than major Western open-weight models. The firm acknowledged Beam outperforms open-weight AI models from Thinking Machines Lab and Nvidia but still trails some Chinese models to an extent. Beam is outranked by Qwen 3.8 Max.
| Efficiency Claim | Comparison Model | Claimed Advantage |
|---|---|---|
| Inference Compute | GLM-5.2 | 3-4x less required |
| General Efficiency | Leading Western Open Models | Over 4x more efficient |
| General Efficiency | GLM-5.2 | 3-4x more efficient |
Training Scale and Methodology
The four-week training regimen for Beam involved a total of 1.3 billion sandboxes. The reinforcement learning process spanned 1 million coding, agentic, and STEM environments and involved 100 million rollouts. Reflection decoupled rollout workers and learner engines in the training pipeline to enhance efficiency.
In a blog post, Reflection noted that early in reinforcement learning, performance improved even as completion lengths fell, as the model learned to solve tasks more effectively with less reasoning. Later, as Beam developed stronger agentic capabilities, completion lengths grew again, with those additional tokens supporting further performance gains. The scale is immense; to match it, DeepSeek would need approximately 16 of its proprietary DSec units, equivalent to roughly 10,500 GB300 GPUs.
Market Context and Strategic Positioning
Beam is positioned as a 'work-ready model' for corporations, public institutions, and developers. Its launch is part of a broader effort by Western AI labs to compete with China's DeepSeek and Zhipu AI in the efficient open-weight model market. While U.S. Firms like OpenAI and Anthropic lead in closed models, they are seen as lagging in the open-weight segment, where model weights are publicly disclosed.
The U.S. Government is actively supporting such initiatives to curb the spread of Chinese AI. Earlier this year, Reflection AI signed a strategic partnership memorandum of understanding with World Group and a sovereign AI factory project, a deal brokered by the U.S. Government. Reflection has also secured substantial funding, raising $2 billion last October and an additional $2.5 billion at a $25 billion pre-money valuation in March. Large-scale computing contracts followed, including a roughly $6.3 billion deal with SpaceX in June and a multibillion-dollar agreement with Navius in July.
Reflection AI plans to release additional technical information on Beam at the end of the month.
