Reflection AI put the weights of Beam, its first open-weight language model, on Hugging Face on 16 October under the Apache 2.0 licence. The model was first introduced about a week earlier. The startup had been known for closed research until now, so the release marks a change of approach, and the company presents the surrounding open stack as the selling point alongside the model's efficiency.
Beam is a sparse mixture of experts with 501 billion parameters, 23 billion of them active per token, built for coding, reasoning and agent work. Weights come in BF16, FP8 and NVFP4, with a standard version and quantised variants. Raw storage is roughly 1TB in BF16 and 500GB in FP8. Reflection says pre-training took under four weeks on 6,144 GB300 NVL72 GPUs, covering 23.8 trillion tokens at 92.3% goodput, while SiliconANGLE describes about two months on several thousand GPUs.
The package goes beyond weights. It includes fine-tuning and post-training recipes, tools for evaluating changes in capability and safety, a drafter model for speculative decoding, serving-engine configurations and fused kernels. Reflection also open-sourced Mirror CLI, the agent harness Beam was trained on. The largest post-training run used more than 100 million reinforcement-learning rollouts on 10,500 GB300 GPUs over four weeks, drawing on about 1.3 billion sandboxes in total.
Support was ready on day one in vLLM, SGLang, Transformers, Unsloth, Axolotl and NVIDIA Dynamo, with more than 40 partners including AWS, Azure, Google Cloud, Databricks, Fireworks and Together, and Dell and HPE for private deployments.
The reliability work is part of the pitch. Reflection built custom algorithms to cope with mismatches between training and inference, plus an automatic rollback when bad data turns up, and says the cluster saw near-zero downtime.
On performance, Reflection says Beam matches the reasoning of GLM-5.2-class models while using three to four times less inference compute, and SiliconANGLE says it comes near models four times its size on some benchmarks. AlphaSignal notes it trails the listed rivals in direct comparisons, and the benchmark and licence claims are the company's own.
About Reflection AI
Reflection AI is a New York-based American company that develops open foundation models and software agents for AI-assisted software development. It was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou.
It came out of stealth in March 2025 with $130 million in financing, and its first product was Asimov, a code-comprehension agent for engineering teams. In October 2025 it raised $2 billion at an $8 billion valuation. In June 2026 it signed a compute agreement with SpaceXAI, which Axios reported would cost $150 million a month, and a month later a deal with Nebius for over $1 billion of computing power. It was valued at $25 billion as of June 2026.
1
Models indexed
Models on AIPopDB
The reports disagree on recipes: AlphaSignal says they shipped with the launch, while SiliconANGLE says self-distillation recipes, which let Beam refine its own output, are due in the coming days. Reflection also plans to share lessons from building Beam with open-source AI tools.
Sources · 2