Reflection AI Introduces Beam: A 501B Open-Weight MoE Model With 23B Active Parameters for Coding and Agentic Workloads

Reflection AI has put Beam forward as its first open-weight model, and it’s not trying to win by brute force. Beam is a sparse mixture-of-experts system with 501B total parameters, but only 23B are active per token. The target is clear: coding, reasoning, and agentic work, with Reflection saying it can compete with larger open models while using 3 to 4 times less inference compute on reasoning benchmarks.

There’s a catch, and it matters. Beam is not ready for self-hosting yet; it’s in final red-teaming, and early access is being handled through a waitlist on Reflection’s own platform. The company also says Kimi K3 still leads on raw capability, so Beam’s main pitch is efficiency at inference time rather than a clean win on every task. Users can tune a reasoning-effort parameter, dialing it down for shorter answers or up for harder problems that need more thought.

The pretraining recipe is unusually aggressive. Beam was trained from scratch on 23.8 trillion tokens drawn from the web, public sources, and proprietary licensed data, with Reflection saying about 95% of raw internet tokens were removed during curation. The team says it preserved roughly 1.8 trillion high-quality tokens that normal filters would have thrown away. It also used an architecture that mixes local and global attention with routed experts, and claims the busiest expert ended pretraining at just 1.04x average load.

The compute story is just as heavy. Pretraining finished in under 4 weeks on 6,144 NVIDIA GB300 NVL72 GPUs, with goodput reaching 92.3% near the end after 9 semi-automatic rewinds. Midtraining pushed the effective context to 1M tokens. Then came the reinforcement-learning phase: 10.5K NVIDIA GB300 GPUs, 4 weeks, more than 100 million rollouts, and about 1.3 billion sandboxes across nearly 1 million coding, agentic, and STEM environments.

Reflection says the RL system stayed stable even with one-day staleness and 107 weight versions of lag. It also reports 110K concurrent rollouts on average, a median of about 12 seconds for new weights to reach inference, and 71 inference incidents handled without stopping training. On its own numbers, Beam scores 80.9 on SWE-bench Verified and 80.1 on Terminal Bench v2.1, which puts it close to the front of the open-weight pack but not at the top everywhere. Apache 2.0 weights are planned for later in October 2026.