Beam: Reflection's 501B Open-weight Model
AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Reflection has introduced Beam, a 501-billion-parameter sparse Mixture-of-Experts model with 23 billion active parameters, aimed at coding, reasoning and agentic tasks. The company says Beam is undergoing final red-teaming and evaluations, with weights and technical materials planned for release later this month; independent verification of its reported benchmark results is not yet available in the supplied material.

Reflection has introduced Beam, a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, designed for coding, reasoning and agentic workloads. The company says it is completing red-teaming and evaluations and plans to release the weights, technical report, model card and developer artifacts later this month, making Beam’s claims open to outside inspection.

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

Reflection describes Beam as its first open-weight model. The company says pretraining used 23.8 trillion tokens drawn from curated web material and proprietary licensed datasets. It says the model matched or outperformed available open base models of similar size, though the announcement does not provide details here that allow readers to independently assess that comparison.

For reinforcement learning, Reflection reports generating more than 100 million rollouts over four weeks on 10,500 NVIDIA GB300 GPUs. It says training and grading used about 1.3 billion sandboxes and that it sourced one million coding, agentic and STEM environments. Reflection calls the effort one of the largest reinforcement-learning runs by an open lab to date; that characterization is the company’s claim.

Reflection says Beam is competitive with larger open models such as GLM 5.2 on coding and agentic tasks and approaches Qwen 3.8-Max on those tasks. It also acknowledges that Kimi K3 remains ahead on raw capability. For advanced reasoning, the company reports scores comparable to GLM 5.2 while using three to four times less inference compute. Its compute estimates use active parameters and generated tokens, and exclude prompt prefill, context-dependent attention operations and serving overhead.

At a glance
announcementWhen: Announced in Reflection’s report; weigh…
The developmentReflection announced Beam as its first open-weight model and said it plans to release its weights and supporting materials later this month.

Lower Compute for Coding Tasks

Beam’s reported results place inference efficiency at the center of its pitch. If the company’s comparisons hold up, a model that performs strongly on coding and agentic work while using fewer active parameters or less generation compute could lower the cost of deploying such systems. That may matter to enterprises running coding assistants or agents repeatedly, where per-task compute can affect operating costs.

The practical value will depend on more than headline parameter counts. Model access, hardware requirements, response quality on real workloads and serving costs will shape whether Beam is useful in production. The planned release of weights and documentation should give developers a way to examine those factors, but the announcement alone does not establish them.

Products Worth Considering

Anker SOLIX C1000 Gen 2 Portable Power Station with 200W Solar Panel

Anker SOLIX C1000 Gen 2 Portable Power Station with 200W Solar Panel

As an affiliate, we earn on qualifying purchases.

Training at Reinforcement Learning Scale

Beam combines large-scale pretraining with reinforcement learning, a method that uses feedback from tasks or environments to refine model behavior. Reflection says it used asynchronous policy-gradient training, where multiple training processes can generate and learn from rollouts at different times. The company reports developing methods to keep learning stable despite stale samples and numerical differences between training and inference systems.

Reflection says some rollouts were more than a day old, corresponding to 107 model weight versions behind the current policy, while training remained stable. It also says Beam was trained with a length penalty intended to reward successful solutions while discouraging unnecessary tokens. These are descriptions of the company’s training approach; the supplied report excerpt does not include the full technical documentation needed to evaluate them.

“Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active, built for coding, reasoning, and agentic workloads.”

— Reflection

Benchmarks Await Independent Review

The supplied announcement reports benchmark comparisons, but it does not provide independent confirmation of Beam’s performance. Reflection says final red-teaming and evaluations are still underway. The planned technical report and model card may add evaluation methods and safety information; until they are available, it is unclear how the results will hold up under outside testing or across different deployments.

Reflection’s inference-compute comparisons are estimates rather than measured serving costs. They exclude prompt prefill, some attention costs and serving overhead, and rely on generated-token counts and active parameters. Actual costs and latency may vary with workloads, hardware and implementation. The announcement also does not specify the final license terms, access conditions or exact release date.

Release Materials Later This Month

Reflection says it plans to publish Beam’s weights, technical report, model card and developer artifacts later this month, after its current red-teaming and evaluation work. The company is accepting sign-ups for early access. The announcement does not give a precise publication date, and it remains unclear whether all materials will be available at once or under what license.

Once the materials are released, developers and researchers will be able to examine the model and test the company’s benchmark and efficiency claims. Those results, along with details on access and deployment, will help clarify Beam’s practical standing among open-weight models.

Key Questions

What is Beam?

Beam is Reflection’s first open-weight model, a sparse Mixture-of-Experts system intended for coding, reasoning and agentic workloads. Reflection reports 501 billion total parameters and 23 billion active parameters.

When will Beam’s weights be released?

Reflection says it plans to release the weights and supporting materials later this month. It has not announced an exact date in the supplied report.

How did Reflection train Beam?

Reflection says it pretrained Beam on 23.8 trillion tokens and generated more than 100 million reinforcement-learning rollouts over four weeks using 10,500 NVIDIA GB300 GPUs.

How does Beam compare with other models?

Reflection says Beam is competitive with GLM 5.2 and approaches Qwen 3.8-Max on coding and agentic tasks, while Kimi K3 remains ahead on raw capability. These are the company’s comparisons and await outside review.

Are Beam’s efficiency claims independently verified?

The supplied announcement does not include independent verification. Reflection describes its compute figures as estimates that exclude prompt prefill, some attention costs and serving overhead.

Source: hn

You May Also Like

14 Ways LegalOn’s Codex Savings Keep Development On Track

LegalOn says it cut Codex costs by half without slowing development, but the cost baseline, timeframe and measurement method are not available.

Google Beam Expands With New Regions, Partners, And Customers

Google Beam broadens its reach with expansion into additional regions, new partners, and a growing customer base, signaling increased adoption of its cloud data transfer services.

Could One Of These 6 AI Writing Assistant Software Picks Suit You?

A ThorstenMeyerAI.com comparison outlines six AI writing tools, their stated uses and the key details buyers should verify before choosing.

What Can Dynamic Literature Context Add To Research Ideation?

A 17-researcher study found LitPivot proposals were rated as better grounded in relevant literature than proposals made with a text editor and AI chat tool.