Two-year-old Brooklyn startup Reflection AI has unveiled Beam, its first frontier open-weight AI model. The company says Beam matches the performance of leading Chinese open models on reasoning benchmarks while using significantly less compute, offering a potential Western alternative in the competitive open-weight AI space.
Reflection AI, founded by two former Google DeepMind researchers, has officially unveiled Beam, its first frontier open-weight AI model. The company says Beam matches the performance of leading Chinese open models such as Z.ai's GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute. The claims have not been independently verified.
Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters. It was pretrained on 23.8 trillion tokens and offers a 1 million token context window. Reflection calls Beam a "workhorse model" aimed at enterprises, the public sector, and developers.
The model enters a competitive field. Reflection is positioning itself against closed labs like OpenAI and Anthropic, as well as Western open-model developers such as Mistral, Meta, and Cohere. Its most direct U.S. competitor is Inkling from Thinking Machines Lab, released in July. Reflection says Beam outperforms Inkling on four coding benchmarks, though Inkling is multimodal while Beam is text-only.
Reflection has raised approximately $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, with a last round valuing the company at $25 billion pre-money. This summer, the startup secured compute deals worth more than $7 billion combined with SpaceX and Nebius, guaranteeing access to Nvidia GB300 chips through 2029.
Looking ahead, Reflection plans to target enterprises and sovereign nations with its "AI factory" product, allowing institutions to build customized local AI systems. The launch of Beam marks a significant step in the race to provide a Western open-weight alternative to the increasingly capable and cost-efficient models emerging from China.
Analysis
Why This Matters
- Beam could provide a cost-effective open-weight alternative to both closed-source models from US labs and increasingly capable open models from China, potentially shifting enterprise AI adoption.
- The model's compute efficiency may lower barriers for organizations wanting to run powerful AI locally rather than relying on API access.
- Reflection's aggressive compute deals and high valuation underscore the capital-intensive nature of frontier AI development.
Background
The open-weight AI model landscape has been dominated in recent months by Chinese labs such as DeepSeek, Qwen, and Z.ai, which have produced strong reasoning models at low cost. Western labs like Meta, Mistral, and Cohere have also released open models, but until recently none matched the performance of the Chinese front-runners. Reflection AI, founded in 2024 by former Google DeepMind researchers, has raised billions to build what it calls "America's open frontier AI lab." Its first model, Beam, is designed to close that gap.
Key Perspectives
Reflection AI: Beam offers a performant, compute-efficient alternative that lets enterprises and sovereign states build customized local AI systems, reducing dependence on proprietary APIs.
Chinese AI labs (DeepSeek, Qwen, Z.ai): They have set a high bar with open models that are both capable and cheap; Beam's claims directly challenge their cost-performance advantage.
Competing Western open model developers (Mistral, Meta, Thinking Machines Lab): Beam raises the stakes for performance in the Western open ecosystem, though it lacks the multimodal capabilities of some rivals.
Critics/Skeptics: Beam's benchmark claims have not been independently verified. It is text-only, which limits applications compared to multimodal models. The cost savings are claimed, not yet proven in real-world deployments.
What to Watch
- Independent benchmark verification or third-party evaluations of Beam's performance.
- Whether enterprises and sovereign nations sign on to use Reflection's "AI factory" product.
- Response from Chinese labs, which may update their models to maintain their cost-performance lead.
