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, a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token, aimed at coding, reasoning and agentic workloads. The company reports large-scale reinforcement learning and competitive benchmark results, but Beam is still undergoing red-teaming and evaluation; weights and technical materials are expected later this month.
Reflection has introduced Beam, its first open-weight model, with 501 billion total parameters and 23 billion active per token. The company is positioning it for coding, reasoning and agentic workloads, and says it plans to publish the model weights and technical materials later this month, after final red-teaming and evaluation.
Beam uses a sparse Mixture-of-Experts architecture, in which only part of the model is active for each token. Reflection says it pretrained Beam on 23.8 trillion tokens from curated web material and proprietary licensed datasets. The company says the model matched or outperformed available open base models of similar size, though its announcement does not provide a full independent assessment of that comparison.
Reflection also describes a large reinforcement-learning effort: more than 100 million rollouts generated over four weeks using 10,500 NVIDIA GB300 GPUs. It says training and grading involved about 1.3 billion sandboxes and one million coding, agentic and STEM environments. Reflection characterizes the run as among the largest conducted by an open lab; that characterization is the company’s claim.
The company reports Beam is competitive with larger open models on some coding and agentic tasks and approaches Qwen 3.8-Max on those tasks. It says Beam is comparable to GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute. Reflection acknowledges that Kimi K3 remains ahead on raw capability. The benchmark comparisons and efficiency estimates are based on the company’s stated evaluation methods and cited external data, rather than a completed independent review of the model.
Lower Compute for Coding Tasks
Beam’s pitch is that a model can deliver useful coding and agentic performance while requiring less inference compute than larger systems. If independent users reproduce Reflection’s estimates, lower compute per task could affect serving costs and make sustained coding-agent use more practical for businesses. The model’s 23 billion active parameters per token, despite 501 billion total, are central to that efficiency claim.
For developers, open weights could allow inspection, adaptation and deployment within their own infrastructure, subject to the eventual license and release terms. Those terms have not yet been included in the announcement. The practical value will depend on reproducible performance, hardware requirements and how the model handles real coding work beyond benchmark tasks.
Products Worth Considering
Jackery HomePower 1000 Plus v2 Portable Power Station
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
As an affiliate, we earn on qualifying purchases.
Training Behind the Announcement
Reflection says Beam’s development combined pretraining with reinforcement learning. Its RL approach used long, multi-step rollouts in coding, STEM and other environments, allowing the model to receive feedback from task outcomes. The company says this scale helped improve performance as the amount of RL compute increased, with no plateau observed in its evaluation suite.
Reflection also describes engineering work to keep asynchronous training stable. In this approach, some training examples can come from model versions that are older than the version currently being updated. The company says it developed methods to manage that policy staleness and reduce differences between training and inference calculations, including learning from interactions generated more than a day earlier.
The announcement includes benchmark tables and estimates that compare model inference compute using active parameter counts and generated tokens. Reflection notes these estimates exclude prompt prefill, context-dependent attention operations and serving overhead. They are approximate comparisons, not direct measurements of total deployment cost.
““Beam’s advantage is efficiency at inference time.””
— Reflection
Release and Results Still Pending
Beam is still undergoing final red-teaming and evaluations, so the announced results may not represent the final release. Reflection has not yet published the weights, technical report, model card or developer artifacts, and the announcement does not specify an exact release date, license or access conditions.
Independent confirmation of the benchmark results and compute comparisons is also pending. The company’s estimates omit several factors that affect real serving costs, and results can vary with hardware, prompts and evaluation setup. It is not yet clear how Beam will perform across broader production use, or how its capabilities and efficiency will compare under consistent independent testing.
Weights Expected Later This Month
Reflection says it plans to release Beam’s weights, technical report, model card and developer artifacts later this month. Interested users can sign up for early access, according to the announcement. The company has not given a specific publication date.
The release materials should clarify the model’s license, deployment requirements and evaluation setup. Independent testing after release will help establish whether the reported coding performance and inference efficiency hold across different hardware and real-world tasks.
Key Questions
What is Beam?
Beam is Reflection’s first open-weight model. It is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active per token, designed for coding, reasoning and agentic workloads.
When will Reflection release Beam’s weights?
Reflection says it plans to release the weights later this month, alongside a technical report, model card and developer artifacts. It has not announced an exact date.
What performance does Reflection report?
Reflection says Beam is competitive with larger open models on some coding and agentic tasks and approaches Qwen 3.8-Max in those areas. It also reports reasoning scores comparable to GLM-5.2 with three to four times less inference compute. Those claims have not yet been independently verified in the announcement.
Is Beam available to use now?
The weights have not yet been released. Reflection says Beam is in final red-teaming and evaluation and offers a sign-up for early access.
Source: hn
