Reflection AI releases Beam open-weight model

Reflection AI has officially launched Beam, its first frontier open-weight AI model, positioning it as a cost-efficient Western alternative to leading Chinese models like DeepSeek and Qwen. According to TechCrunch, the Brooklyn-based startup claims Beam matches the performance of top Chinese open models on advanced reasoning benchmarks while operating at a fraction of the token cost and inference time. This release confirms earlier reporting from Axios that the company was close to a launch, with Reflection sharing detailed specifications in a Monday blog post.

Technical Specifications and Performance

Beam is a text-only mixture-of-experts (MoE) architecture with 501 billion total parameters, of which 23 billion are active during inference. The model was pretrained on 23.8 trillion tokens and supports a 1 million token context window. Reflection states that Beam was trained using high-compute reinforcement learning to optimize for reasoning, coding, and agentic tasks.

To put the scale in perspective, Z.ai’s GLM-5.2, a direct competitor, features roughly 744 billion total parameters with 40 billion active. Despite having fewer active parameters, Reflection claims Beam scores on par with GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute. The company describes Beam as a "workhorse model" designed for enterprises, the public sector, and developers who need high-performance reasoning without the associated compute overhead.

Competitive Landscape

Reflection is positioning Beam against three distinct groups: closed labs like Anthropic and OpenAI, Chinese open-weight models, and Western open-model providers like Mistral, Meta, and Cohere. Its most direct U.S. rival is likely Inkling, the open model released by Mira Murati’s Thinking Machines Lab in July.

Reflection’s internal benchmarks show that Beam outscores Inkling on four coding tests where both models report results. However, a key differentiator remains that Inkling is a multimodal model, whereas Beam is strictly text-only. This limitation may restrict Beam’s utility in applications requiring visual or audio processing, but its focus on text efficiency aligns with the needs of coding and agentic workflows.

Compute Strategy and Funding

Reflection was founded in 2024 by two former Google DeepMind researchers and has raised approximately $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners. The company’s last round valued it at $25 billion pre-money.

A critical part of Reflection’s strategy involves securing massive compute resources. This summer, the startup signed deals worth more than $7 billion with SpaceX and Nebius to secure access to Nvidia’s GB300 chips through 2029. This infrastructure lock-in is essential for training frontier models capable of competing with both closed Western models and cheaper Chinese open-weight alternatives.

The AI Factory Vision

Reflection is targeting enterprises and sovereign nations with a product concept called "AI factories." This approach allows institutions to build customized, local AI systems by training Reflection’s models on their own proprietary data. Nvidia CEO Jensen Huang, a major backer of Reflection, has championed this vision, noting that it strengthens the open AI ecosystem while benefiting Nvidia’s GPU sales.

Axios reported that hedge funds and trading firms are eager to adopt such systems. Reflection has already begun testing this concept with Shinsegae Group in South Korea, aiming to establish a sovereign AI factory partnership. This move signals a shift toward localized, data-sovereign AI deployments, particularly in sectors where data privacy and regulatory compliance are paramount.

What to Watch

• Independent Verification: Reflection’s performance claims have not yet been independently verified. Third-party benchmarks will be crucial to confirm whether Beam truly matches GLM-5.2 on reasoning tasks.
• Weight Release: The company plans to release Beam’s weights and full technical details this month, with distribution through hyperscalers and neoclouds.
• Integration Ecosystem: Launch integrations across open-source libraries will determine how easily developers can adopt Beam into existing workflows.
• Sovereign Deals: The success of the Shinsegae partnership could serve as a template for other nations seeking to build local AI capabilities.