Reflection AI Launches Beam a 501B Open-Weight Model Built for Coding Agents

US-based Reflection AI has introduced its first open-weight model, called Beam. It is a sparse Mixture of Experts (MoE) — a design where each piece of input only uses a small subset of the model's total parameters — with 501 billion total parameters but only 23 billion active per token ¹². The model is aimed at coding, reasoning, and agentic workloads, which means it can use tools and take multi-step actions to complete tasks.

Performance and Competition

Reflection says Beam directly competes with much larger open models while using three to four times less inference compute on reasoning benchmarks ¹. The company's leader board shows Beam scoring 80.9 on SWE-bench Verified, compared to NVIDIA's Nemotron 3 Ultra at 70.7, and 80.1 on Terminal Bench v2.1, close to Z.ai's GLM 5.2 at 81.0 ². Frontier models like Kimi K3 from Moonshot AI and DeepSeek V4.1 Flash still lead those same benchmarks at 88.3 and 90.6 respectively .

Reflection is upfront about the gap. Where Kimi K3 remains ahead on raw capability, Beam's pitch is inference-time efficiency ². Users can control how much reasoning effort the model expends with a simple parameter — lower settings give shorter answers, higher settings allow deeper reasoning on harder tasks ¹.

Training at Scale

Beam was pretrained from scratch on 23.8 trillion tokens scraped from the web and proprietary licensed datasets ¹². The Reflection team says its curation pipeline removed about 95 percent of raw internet tokens while keeping roughly 1.8 trillion high-quality tokens that conventional filters would have discarded . Pretraining ran on 6,144 NVIDIA GB300 NVL72 GPUs in under four weeks, reaching 92.3 percent goodput near the end .

Reinforcement learning formed the central scaling axis. The RL run used 10,500 GB300 GPUs over four weeks and generated over 100 million rollouts across nearly one million coding, agentic, and STEM training environments ¹². The system sustained 110,000 concurrent rollouts on average, and new weight updates reached the inference fleet in about 12 seconds .

Availability

Beam is not yet available for self-hosted deployment. It is currently in final red-teaming and safety evaluation ¹². Early access runs through a waitlist on the Reflection AI platform . Full weights under an Apache 2.0 license, along with a technical report and model card, are planned for release later in October 2026 . Reflection also trained a separate safety and alignment teacher model merged with the RL teacher through multi-teacher on-policy distillation, though full safety results will appear in the technical report .

Key facts

Beam is a 501B total-parameter MoE model with 23B active parameters per token
Scores 80.9 on SWE-bench Verified and 80.1 on Terminal Bench v2.1
Pretrained on 23.8 trillion tokens using 6,144 NVIDIA GB300 GPUs in under 4 weeks
RL training generated 100M+ rollouts on 10,500 GB300 GPUs over 4 weeks
Open weights planned for October 2026 under Apache 2.0 license

Sources

Reflection AI Introduces Beam: A 501B Open-Weight MoE Model With 23B Active Parameters for Coding and Agentic Workloads — marktechpost.com
Introducing Beam: Reflection’s 501B open-weight model - alphaXiv