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 tasks. The company reports large-scale pretraining and reinforcement learning, but Beam remains in final red-teaming and evaluation; its weights and technical materials have not yet been released.
Reflection has introduced Beam, its first open-weight model, with 501 billion total parameters and 23 billion active per token. The company says the model is designed for coding, reasoning and agentic tasks; its weights are not yet available, as Beam remains in final red-teaming and evaluations.
Reflection describes Beam as a sparse Mixture-of-Experts model. The company says it pretrained the model on 23.8 trillion tokens drawn from curated web material and proprietary licensed datasets. It also reports that the model’s capabilities were developed through reinforcement learning, or RL, alongside pretraining.
For the RL campaign, Reflection says it used 10,500 NVIDIA GB300 GPUs over four weeks, generating more than 100 million rollouts. The company also reports using approximately 1.3 billion sandboxes for training and grading and sourcing one million coding, agentic and STEM environments. These are figures reported by Reflection; its technical report and other supporting materials have not yet been published.
Reflection says Beam is competitive with some larger open models on coding and agentic tasks and approaches Qwen 3.8-Max on those tasks. The company also says Beam’s advanced-reasoning scores are comparable to GLM-5.2 while using three to four times less inference compute. Those are company-reported comparisons, not an independent assessment in the announcement. Reflection says it will publish benchmark information and methodology with the planned release materials.
A Smaller Active Model for Coding
Beam’s headline size can obscure the distinction between total and active parameters: although the model has 501 billion parameters in total, Reflection says 23 billion are active per token. That design and the company’s efficiency claims matter to developers weighing model capability against the compute needed to serve it. If the reported results hold up under scrutiny, Beam could offer a comparatively efficient option for coding and agentic workloads.
For businesses, the practical questions are not limited to benchmark scores. They include the cost and reliability of running the model, how well it handles multi-step work and tool use, and whether the eventual license permits their intended use. Open weights can give developers more control than a hosted-only service, but actual access terms and deployment requirements depend on what Reflection releases. Those details are not yet available.
As an affiliate, we earn on qualifying purchases.
Reflection’s First Open-Weight Release
Reflection’s announcement positions Beam as the company’s first open-weight model, with a particular focus on coding and agentic performance. The company says it invested in both large-scale pretraining and RL rather than relying on pretraining alone. Its account describes rollouts—model-generated attempts or interactions used during RL—as a way to train multi-step problem-solving, tool use and responses to feedback from environments.
Reflection also says it used asynchronous policy-gradient training and developed methods to address stale training samples and numerical mismatches between training and inference systems. It reports stable learning even when some training interactions were more than a day old and 107 weight versions behind the current policy. The announcement presents these as details of its training approach; independent technical review is pending.
The company’s benchmark comparisons include coding, terminal-use, reasoning and STEM evaluations, with some entries marked as unreported. It says its compute estimates use active rather than total parameters for mixture-of-experts models and exclude prompt prefill, some attention operations and serving overhead. That caveat means the figures are estimates for comparing generation compute, not direct measurements of end-to-end serving cost.
“Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active, built for coding, reasoning, and agentic workloads.”
— Reflection
As an affiliate, we earn on qualifying purchases.
Independent Results Still Pending
Beam’s weights, model card and technical report are not yet public, so outside researchers cannot yet inspect the model or reproduce the company’s training and benchmark claims from those materials. The announcement does not provide enough information to independently verify its comparisons with named models, including the evaluation setup, model versions and full results across tasks.
It is also unclear what license will govern the weights, what hardware and serving setup users will need, or how Reflection’s estimated inference-compute advantage translates into real-world latency and cost. The company says its calculations exclude some serving-related work, so they should not be read as a complete measure of deployment expense. Red-teaming findings and any resulting changes have not been described.
large scale AI model deployment tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Release Materials Due Later This Month
Reflection says it plans to publish Beam’s weights, technical report, model card and developer artifacts later this month, after final red-teaming and evaluations. The company is offering early access through sign-up, but the announcement does not specify who will qualify or when access will begin.
Once the materials are available, developers and independent evaluators will be able to examine the license, test the model on their own workloads and assess the reported benchmark and efficiency claims. The release date, access conditions and final evaluation results remain subject to Reflection’s forthcoming updates.
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 large is Beam?
Reflection reports 501 billion total parameters, with 23 billion active per token. Those figures describe total and active model parameters, respectively.
Can developers download Beam now?
No. Reflection says Beam is still undergoing final red-teaming and evaluations. The company plans to release weights and supporting materials later this month and says users can sign up for early access.
How does Reflection say Beam performs?
Reflection says Beam is competitive with larger open models on some coding and agentic tasks and reports advanced-reasoning performance comparable to GLM-5.2 with three to four times less inference compute. These are company-reported claims; independent verification and the full technical report are pending.
What details are not yet known?
The announcement does not specify the final license, access terms, deployment requirements or independent evaluation results. Those details are expected to become clearer when Reflection publishes the weights and documentation.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
