Quick Facts
- Beam uses 501 billion total parameters but only 23 billion active parameters per token, which Reflection claims cuts inference compute by three to four times versus comparable open models.
- Model weights shipped October 16 on Hugging Face under an Apache 2.0 license; independent benchmarker Artificial Analysis ranked Beam second among U.S. open-weight models with a score of 30 on its Intelligence Index.
- Reflection trained Beam on 23.8 trillion tokens in under four weeks using 6,144 Nvidia GB300 GPUs, and ran over 100 million reinforcement learning rollouts on 10,500 GB300 GPUs.
Reflection AI released Beam on October 16, making the weights of its debut open-weight model available to developers and enterprises after announcing the model on October 5. The company argues that the cost of a finished task matters more than the cost per token, and that Beam's architecture delivers on that bet.
The model uses a sparse mixture-of-experts design with 501 billion total parameters and 23 billion active parameters. Because active parameters determine compute per token, each Beam token costs roughly half the compute of a token from Meta's Muse Spark Open, which activates 42 billion parameters. The context window stretches to one million tokens.
A Penalty for Wasted Tokens
The core engineering decision behind Beam is a reinforcement learning technique that penalizes the model for using unnecessary tokens. Ioannis Antonoglou, co-founder and CTO, described the challenge: a poorly tuned penalty teaches the model to quit rather than be concise. With enough reinforcement learning, the model learns to solve the task and stay brief at the same time.
The training run consumed serious infrastructure. Reflection pre-trained Beam on 23.8 trillion tokens across 6,144 Nvidia GB300 GPUs in under four weeks. Reinforcement learning added more than 100 million rollouts on 10,500 GB300 GPUs over another four weeks. The company ran 1.3 billion sandboxes in total, with its compute fleet peaking at 883,006 virtual CPUs spread across 23 clusters on Google Cloud and Amazon Web Services.
Benchmark Results and Gaps
Reflection published scores of 80.9 on SWEBench Verified, 97.8 on AIME 2026, 90.5 on GPQA Diamond, and 36.2 on Humanity's Last Exam without tools. These figures are vendor-reported and had not been independently verified at the time of release.
Independent testing by Artificial Analysis placed Beam at 30 on its Intelligence Index, a composite of ten evaluations covering agentic work, coding, science, and general knowledge. That ranked Beam second among open-weight models from U.S. labs, behind Meta's Muse Spark Open at 39 and ahead of Nvidia's Nemotron 3.5 Super VL at 26. Artificial Analysis also confirmed the token efficiency claim, finding Beam used roughly 26,000 output tokens per Intelligence Index task, less than half the approximately 52,000 tokens K2 Horizon used to score just one point higher.
Gaps remain against top Chinese models. On DeepSWE, Beam scored 44.4 versus Kimi's 68.0 and DeepSeek's 74.2. Beam's hallucination rate also stands out: Artificial Analysis measured it at 47%, compared with 26% for K2 Horizon and 30% for Qwen3.8 27B.
Business Model and Release Terms
The weights are available on Hugging Face under an Apache 2.0 license in BF16, FP8, and NVFP4 formats. Reflection also released a 1.93 billion-parameter speculative decoding drafter called DFlash2 alongside the main model.
CEO Misha Laskin described the open-weight strategy in direct terms: the only way to own intelligence is, by definition, if it is open. Researchers can use the model at no cost. Revenue will come from large enterprises building products on top of Reflection's models and from governments developing sovereign AI systems.
Antonoglou said enterprises and governments want full control over their AI, which positions Beam squarely against cloud-only proprietary models. The release comes days after the Financial Times reported that Nvidia was in early talks to acquire Reflection AI or deepen its investment in the company, though no deal has been announced.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.