TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Reflection has introduced Beam, its first open-weight model, a sparse Mixture-of-Experts system with 501 billion total parameters and 23 billion active. The company reports strong coding and agentic benchmark results and says Beam uses less inference compute than some larger models; those comparisons are company-reported, and the model is still undergoing final red-teaming and evaluation. Reflection says it plans to publish the weights and technical materials later this month.
Reflection has introduced Beam, its first open-weight model, describing it as a 501-billion-parameter sparse Mixture-of-Experts system with 23 billion parameters active at a time. The company says the model is aimed at coding, reasoning and agentic workloads; its weights, technical report and other developer materials are not yet public, and Beam remains in final red-teaming and evaluation.
Reflection says Beam was pretrained on 23.8 trillion curated tokens from web and proprietary licensed datasets. The company reports that its high-compute reinforcement-learning campaign generated more than 100 million rollouts using 10,500 NVIDIA GB300 GPUs over four weeks. It also says training and grading used approximately 1.3 billion sandboxes and drew on one million coding, agentic and STEM environments.
In its announcement, Reflection said Beam was competitive with larger open models on coding and agentic tasks, and that its reasoning benchmark scores were comparable to GLM-5.2 while using three to four times less inference compute. The company also said some models, including Kimi K3, remain ahead on raw capability, while Beam’s stated advantage is inference efficiency. These are vendor-reported comparisons, not an independent evaluation presented in the announcement.
Reflection said it will offer early access through sign-up and plans to release the model weights, model card, technical report and developer artifacts later this month. The source announcement does not provide a specific calendar date for that release.
Lower Compute for Coding Workloads
Beam’s importance will depend not only on its benchmark scores but on whether developers can run it reliably and affordably. A model with 23 billion active parameters despite 501 billion total parameters may require less computation per generated token than a dense model of similar total size. Reflection presents that efficiency as a potential advantage for enterprise coding and agentic applications, where systems may need to reason through tasks and use tools over multiple steps.
However, efficiency claims need to be read with their measurement limits. Reflection’s published estimate of generation compute uses active parameters and generated tokens, and excludes prompt prefill, context-dependent attention operations and serving overhead. It is an approximate comparison, not a direct measurement of a customer’s total operating costs. Access to the weights and technical documentation will let developers and independent evaluators examine the claims in more detail.
As an affiliate, we earn on qualifying purchases.
How Reflection Built Beam
Beam is the company’s first open-weight release, according to Reflection. Its announcement emphasizes pretraining and reinforcement learning as complementary parts of development: a large token corpus supplied the initial training, while the RL campaign trained the model through repeated interactions with tasks and environments.
Reflection says it used asynchronous policy-gradient methods and developed techniques to address stale training data and differences between training and inference systems. It reports that learning remained stable even with samples more than a day old and 107 model-weight versions behind the current policy. The company also describes a controllable length penalty intended to reward successful answers while discouraging unnecessary tokens. These details are Reflection’s account; the promised technical report should provide more information about methods and evaluation.
““Beam advances the Western open-weight frontier and is competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks.””
— Reflection, in its Beam announcement
As an affiliate, we earn on qualifying purchases.
Independent Testing Still Pending
Beam’s weights and full technical report were not available in the supplied announcement, so independent testing cannot yet confirm its benchmark results, compute efficiency or reproducibility. The company’s comparisons also rely on evaluations from multiple sources, and its compute estimates omit some costs of actual deployment. It is not yet clear what access conditions, licensing terms or hardware requirements will apply when the model is released. Reflection has not specified an exact release date beyond saying the materials are planned for later this month.
open-weight machine learning models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Weights and Evaluations Expected
Reflection says Beam is undergoing final red-teaming and evaluations, with early access available by sign-up. The next substantive milestone is the planned publication of the weights, model card, technical report and developer artifacts later this month. Those materials should clarify how the model can be used, what safety testing has been completed and how its reported results were measured. Independent evaluations and reports from developers will be needed to establish how Beam performs outside the company’s own testing.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Beam?
Beam is Reflection’s first open-weight model, described by the company as a sparse Mixture-of-Experts system for coding, reasoning and agentic tasks.
How many parameters does Beam have?
Reflection reports 501 billion total parameters, with 23 billion active at a time.
Are Beam’s weights available now?
Not according to the announcement. Reflection says the model is undergoing final red-teaming and evaluations and plans to release the weights and supporting materials later this month.
Are Beam’s performance claims independently verified?
The benchmark and efficiency comparisons in the announcement are Reflection’s reported results. Independent confirmation will require access to the model and supporting evaluation details.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
