Beam: Reflection Launches an Open 501B MoE Betting on Massive RL

A workhorse model rather than a champion

Reflection presents Beam as its first open-weight model, built for code, reasoning and agentic workloads. The architecture is a Mixture-of-Experts: 501 billion parameters total, but only 23 billion engaged per token.

The positioning is deliberate. Beam claims to be competitive with GLM 5.2 and close to Qwen 3.8-Max on code and agentic tasks, while acknowledging Kimi K3 remains ahead in raw capability. The pitch is inference efficiency.

The published numbers confirm this nuanced picture: 80.9 on SWE-bench Verified, 77.2 on SWE Bench Pro v2-Hard, but 44.4 on DeepSWE versus 68.0 for Kimi K3 and 74.2 for DeepSeek V4.1 Flash. The efficiency comparison relies on an estimate (2 × active parameters × generated tokens) that excludes prefill and serving costs.

RL as the main scaling axis

The core of the post focuses on reinforcement learning. Reflection mobilized 10,500 GB300 GPUs for four weeks to generate over 100 million rollouts, with contexts up to 256K tokens. For comparison, the lab notes that Inkling was trained on 30 million rollouts and MiMo on 753,000.

The lab claims scores on Terminal-Bench, HLE and DeepSWE kept improving with RL compute, with no visible plateau. It used fully asynchronous policy gradients, with new algorithms to stay stable even when learning from interactions over a day old, i.e. 107 weight versions behind.

A controllable length penalty first taught the model to solve tasks with fewer tokens, before agentic capabilities justified longer responses. The user retains control via a reasoning effort parameter.

Generalization and environments

Trained on reasoning, software engineering and terminal tasks, Beam improved at web navigation despite never having browsing tasks. Given web access, it learned on its own to query other LLMs and use OCR APIs to read documents.

This RL relies on nearly a million environments, mostly synthetic, supplemented by vendor and open source data. They were filtered to be neither trivial nor impossible, nor hackable. According to Reflection, every compromise on task quality led to capability plateaus.

Infrastructure built in-house

The RL platform supported an average of 110,000 concurrent rollouts and up to 170,000 concurrent sandboxes, across more than 20 clusters, two clouds and four regions. New weights reached the inference fleet in about 12 seconds median, and 71 inference incidents were absorbed without halting training.

Pre-training ran in under four weeks on 6,144 GB300 NVL72 GPUs, with an in-house Kubernetes scheduler and a system for detecting silent data corruption. Result: only nine semi-automatic rewinds and 92.3% goodput by the end of the run.

Data, architecture and alignment

The 23.8 trillion tokens come from the web, public sources and proprietary licensed datasets. About 95% of raw tokens are discarded, but Reflection claims to recover 1.8 trillion quality tokens that standard filters would have dropped, 87% of which from its curated web code.

The architecture combines interleaved local and global attention, finely routed experts and several load balancing mechanisms inspired by DeepSeek. The most loaded expert never exceeds 1.04 times the average by the end of pre-training, across 52 layers.

On safety, a second model was trained specifically on alignment, then merged with the RL model via on-policy multi-teacher distillation. Reflection promises to publish its safety evaluations and open-source its tools.