Reflection AI open-sources Beam model weights and post-training recipes

On October 16, 2026, AI startup Reflection AI released the open-source weights and post-training recipes for Beam, its 501-billion-parameter open model licensed under Apache 2.0. Trained using 10,500 Nvidia GB300 GPUs across four weeks of intensive reinforcement learning, Beam penalizes unnecessary token output to drastically cut task costs while competing against Chinese open-weight models like DeepSeek V4.1 Flash.

Release of the Beam model

• ▪Reflection AI previewed the details of Beam in an October 5, 2026 blog post prior to releasing Beam's open-source weights on October 16, 2026
• ▪Reflection AI released post-training recipes, code templates, and safety measurement tools alongside Beam's weights on October 16, 2026, to enable developers to fine-tune the Beam model
• ▪Reflection AI Inc. released the open-source weights of its 501-billion-parameter Beam model on Hugging Face under the Apache 2.0 license on October 16, 2026

Token efficiency in training and benchmarks

• ▪On the Terminal-Bench 2.1 benchmark, Reflection AI's Beam model achieved a score of approximately 80% while generating roughly 25,000 tokens per attempt, compared to Z.ai's GLM-5.2 generating roughly 80,000 tokens
• ▪Reflection AI trained Beam using intensive reinforcement learning that penalized the model whenever it generated unnecessary tokens to complete a task
• ▪On the DeepSWE software engineering benchmark, Reflection AI's Beam model generated a little more than half as many tokens as Z.ai's GLM-5.2 while achieving approximately the same score

Training Infrastructure and reliability

• ▪Reflection AI conducted a four-week reinforcement learning run for its Beam model using a cluster of 10,500 Nvidia GB300 GPUs and executing over 100 million rollouts
• ▪Reflection AI trained its Beam model over two months using several thousand graphics cards with virtually zero cluster downtime
• ▪During training and evaluation of Reflection AI's Beam model, Reflection AI operated up to 1.3 billion sandboxes across 23 Google Cloud and Amazon Web Services clusters, peaking at 883,006 virtual CPUs
• ▪Reflection AI developed automated reliability optimization algorithms that detect data quality errors and revert training runs to earlier stages instead of restarting from scratch

Inference and deployment optimizations

• ▪Reflection AI released a lightweight drafter model alongside Beam to enable speculative decoding and reduce hardware usage during inference
• ▪The 8-bit and 4-bit quantized versions of Reflection AI's Beam model can run on an eight-GPU server powered by Nvidia H100, H200, or GB300 chips
• ▪Reflection AI provided pre-configured Beam integrations for popular serving engines and custom fused kernels to speed up processing
• ▪Open-source serving software vLLM announced plans to publish demonstration code showing Reflection AI's Beam model running in the 4-bit MXFP4 format supported by AMD and Nvidia Blackwell chips

Enterprise deployment and tools

• ▪Reflection AI released Mirror CLI, an open-source coding agent written in Rust and Python, alongside an open operating system built with 33 named launch partners
• ▪Reflection AI partnered with South Korean retail conglomerate Shinsegae Group to construct an AI factory and develop shopping agents using its Beam model

Competition with Chinese open-weight models

• ▪Reflection AI designed Beam to compete against open-weight Chinese AI models, including DeepSeek V4.1 Flash, Z.ai's GLM-5.2, and Alibaba's Qwen 3.8
• ▪In a September 2026 VentureBeat survey of 32 organizations running open-weight models, 56% of respondents could not state with confidence whether any models in their tech stack originated from a Chinese lab
• ▪U.S. companies including DoorDash, Coinbase, Pinterest, and Airbnb reported reducing AI costs by deploying open-weight models such as Alibaba's Qwen and Moonshot's Kimi
• ▪On the Terminal-Bench 2.1 benchmark, DeepSeek V4.1 Flash scored 90.6% compared to Reflection AI's Beam model scoring 80.1%