Beam: Reflection's 501B open-weight model

Model inference costs fall significantly for agentic coding tasks.

Reflection has introduced Beam, its first open-weight model. Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, of which 23 billion are active, designed for coding, reasoning, and agentic tasks. The model was pretrained on 23.8 trillion tokens from web and proprietary datasets, reportedly matching or exceeding similar-sized open models. Reflection also invested heavily in high-compute reinforcement learning (RL), using 10.5K NVIDIA GB300 GPUs for four weeks to generate over 100 million rollouts. This RL phase focused on improving problem-solving strategies and multi-step reasoning. Beam demonstrates competitive performance on coding and agentic benchmarks, approaching models like Qwen 3.8-Max, with a key advantage in inference efficiency. Reflection is releasing the model weights, technical report, and associated artifacts later this month after final safety evaluations.

Beam's architecture and training emphasize inference efficiency, achieving comparable performance to larger models with substantially less computational cost. This is particularly relevant for enterprise applications demanding continuous operation and rapid response times. The model's efficient reasoning capabilities, coupled with its focus on coding and agentic workloads, position it as a strong contender for businesses looking to integrate advanced AI into their workflows without incurring prohibitive compute expenses.

The large-scale reinforcement learning investment, utilizing over 100 million rollouts, indicates a significant shift towards developing more robust and generalizable agentic AI. By training on diverse environments and focusing on multi-step reasoning and tool use, Reflection is pushing the boundaries of what open-weight models can achieve in autonomous task completion. This approach suggests future open models will increasingly leverage extensive RL for emergent capabilities, potentially democratizing access to sophisticated AI agents.