Reflection unveils Beam, a 501B-parameter open-weight model
Beam is a mixture-of-experts model that activates only 23 billion parameters per step. It posts strong coding and agent scores, and its weights ship later in October under Apache 2.0.
What happened
On October 5 Reflection AI introduced Beam, its first open-weight model: a sparse mixture-of-experts system with 501 billion total parameters, of which 23 billion are active per step, built for coding, reasoning and agentic work. Weights and technical documentation are due later in October under the Apache 2.0 license.
The numbers
Reflection reports 80.1 on Terminal-Bench v2.1, 80.9 on SWE-Bench Verified, 77.2 on SWE-Bench Pro v2-Hard, 97.8 on AIME 2026, 90.5 on GPQA Diamond and 36.2 on Humanity's Last Exam without tools. It says Beam matches GLM-5.2 on advanced reasoning while using three to four times less inference compute.
Training
Beam was pretrained on 23.8 trillion tokens from the web, public sources and licensed proprietary data, in under four weeks on 6,144 Nvidia GB300 NVL72 GPUs, then tuned with reinforcement learning for coding and agent tasks.
Open weights and mixture of experts
An open-weight model publishes its trained parameters so anyone can run it on their own servers and fine-tune it. A mixture-of-experts model holds many sub-networks and activates only a few per input, so Beam carries the knowledge of a large model while spending compute like a mid-sized one.
Caveats
The scores are Reflection's own and independent tests will follow the weight release. Even with 23 billion active parameters, all 501 billion must sit in memory, so most users will reach Beam through hosted endpoints rather than their own servers.
Why it matters
Chinese labs such as DeepSeek, Qwen and GLM have led open-weight models over the past year. A US model at this scale under a permissive license gives companies that want to keep data in-house another option.
Who Reflection is
Reflection AI is a US startup focused on open models for coding and agents. Beam is its first open-weight release, and its Apache 2.0 license lets companies use, modify and sell products built on it with few conditions, which matters for firms that cannot send data to closed APIs.
What it means for your business
For work involving customer data, contracts or source code, open-weight models let data stay on your own infrastructure. Judge Beam on your real tasks once the weights are out.