Reflection Introduces Beam, a Planned Open-Weight Model
Reflection has introduced Beam, a text-based sparse Mixture-of-Experts model. Its weights are not yet publicly available; the company plans to release them under the Apache 2.0 license later in October 2026.
On October 5, Reflection introduced Beam, a planned open-weight model. Beam is currently undergoing final red-teaming and evaluation, so it is not publicly available for download. At present, access is limited to a selected group of users through a waitlist.
The company plans to release the weights under the Apache 2.0 license later in October 2026. Along with them, it is expected to release documentation and tools for running, evaluating, and fine-tuning the model. Until then, Beam does not have publicly available weights.
What the planned open-weight model offers
Beam is a text model built on a sparse Mixture-of-Experts (MoE) architecture. It has 501 billion total parameters and 23 billion active parameters.
Reflection positions Beam for programming, reasoning, and agent workloads. The model has a context window of one million tokens. The company also states that pretraining was conducted on 23.8 trillion tokens.
According to the company, RL training involved more than 100 million rollouts. The process reportedly used 10,500 NVIDIA GB300 GPUs over four weeks.
Performance comparisons have not yet been independently confirmed
Reflection claims that Beam achieves comparable reasoning performance to the Chinese GLM-5.2 model, while reportedly requiring approximately three to four times less inference compute. This claim has not been independently verified.
The company itself cautions that its efficiency comparison is an indicative calculation based on FLOPS and tokens. It does not include prompt prefill, context-length-dependent attention operations, or operational overhead. The comparison therefore cannot yet be used to infer specific operating costs in real-world deployments.
Independent benchmarks, inference measurements, and published technical materials will be important for assessing the model. The results of the ongoing red-teaming and safety evaluations will also be relevant.
What comes next
If Reflection follows through on its plan, it will make Beam’s weights available under the Apache 2.0 license during October, along with documentation and tools for running, evaluating, and fine-tuning the model.
The next things to watch will be whether the company delivers the promised weights on the announced timeline, what the safety testing results will be, and to what extent the model will be available through distribution partners. Only after the weights are released will it be possible to independently verify the stated performance and practical operating requirements.
Sources
- Reflection — Introducing Beam: Reflection’s 501B open-weight model – Primary confirmation of Beam’s introduction, parameters, training data, red-teaming status, waitlist, and planned release of the weights under Apache 2.0.
- TechCrunch — Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost – Independent confirmation that the claims about comparable performance and lower compute requirements have not been independently verified, and that the weights are not expected until later in October.
Verified and updated: 10/06/2026 06:26
