AI startup Reflection has introduced Beam, a 501 billion-parameter mixture-of-experts (MoE) language model with a 23 billion active parameter subset. The company markets Beam chiefly for its token efficiency — claiming it can perform intelligent reasoning while consuming significantly fewer tokens than many comparable models.
Basis for efficiency claims
Reflection says Beam’s pretraining and reinforcement-learning fine-tuning involved substantial compute. The pretraining phase ran for four weeks on 6,144 GB300 NVL72 GPUs, followed by an RL run of roughly four weeks using about 10,500 of the same GPUs. According to the company, the pretraining dataset comprised “23.8 trillion diverse, curated, high-quality tokens” drawn from web sources and proprietary licensed datasets; Reflection claims this matches or outperforms similarly sized open base models.
The team emphasizes that the data pipeline was designed to give Beam a foundation for downstream agentic coding tasks — source code, technical explanations, and mathematical and scientific knowledge preserved through curation.
Published benchmark scores
Reflection has published several benchmark results that place Beam behind some leading Chinese models:
- DeepSWE 1.1: Beam 44.4 vs. GLM-5.3 61
- GPQA Diamond: Beam 90.5 vs. GLM-5.3 91.7
- MCP Atlas: Beam 78.7 vs. GLM-5.3 84.2
These figures indicate that while Beam is competitive on certain measures, it trails frontier models from companies such as Zhipu (GLM), Moonshot (Kimi), and DeepSeek on the reported metrics.
Token efficiency and RL training details
Reflection states that Beam combines coding and agentic capabilities with highly efficient reasoning: on advanced reasoning benchmarks it reportedly achieves scores comparable to GLM-5.2 while using 3–4× less inference compute. Part of this efficiency, the company attributes, comes from careful RL training: they say they sourced one million high-quality coding, agentic, and STEM environments for fine-tuning, and claim this is one of the largest-scale RL runs performed by any open lab to date.
Why this matters — a potential US return to open-weight models
Reflection positions Beam as a possible step toward bringing the US back into the open-weight model frontier. Aside from Google’s smaller Gemma-class models, the US has lagged on open-weight frontier models — an effect Reflection attributes in part to Meta stepping back from releasing competitive LLaMA variants and to Meta’s MSL team still ramping up. Models and companies such as Beam and Reflection could provide a route for US-based open-weight efforts to re-enter the leading tier.
Weights and documentation release
Reflection wrote at the beginning of October that they will release the weights, a technical report, a model card, and developer artifacts later that month. The planned open release would allow external researchers to perform independent evaluations and comparisons.
Summary
Beam is a token-efficient 501B-parameter MoE model trained with large compute and a curated 23.8 trillion-token dataset. While the company emphasizes efficiency and agentic coding capabilities supported by large-scale RL, the published benchmark numbers show Beam currently lags the capability frontier held by some Chinese LLMs. The forthcoming public release of weights and documentation will enable independent verification.
