Reflection unveils Beam, 501 billion parameters, targeting businesses

Reflection AI presents Beam, an open-weight model with 501 billion parameters, of which 23 billion are active, claiming performance close to the best Chinese models with 3 to 4 times less inference computation. With computing contracts exceeding $7 billion and estimated funding of $4.7 billion, the company targets enterprises and sovereign nations with its concept of an "AI factory." Comparisons remain independently unverified.

Billions for Computing and a Valuation of $25 Billion

Reflection AI has secured large-scale computing capabilities, notably through contracts worth over $7 billion signed this summer with SpaceX and Nebius to ensure access to Nvidia GB300 chips until 2029. The company claims to have the necessary infrastructure for training cutting-edge models. According to PitchBook, Reflection AI has raised approximately $4.7 billion from investors like Nvidia, Sequoia Capital, and Lightspeed Venture Partners, and its latest funding round values it at $25 billion pre-investment. Jensen Huang, CEO of Nvidia, supports the "AI factory" vision, and Nvidia provides the GPUs that would power these systems.

Beam: Textual Mixture of Experts and XXL Specifications

Reflection AI has unveiled Beam, a mixture-of-experts model dedicated to text, trained through reinforcement learning requiring significant computing power. Beam incorporates 501 billion parameters, of which 23 billion are used during inference, has been pre-trained on 23.8 trillion tokens, and offers a context window of 1 million tokens. Reflection AI highlights applications in reasoning, coding, and agentic tasks, while promising a reduction in token costs and inference time compared to its competitors.

Claimed Performance Against Chinese and Western Models

The performance claims made by Reflection AI for Beam have not been independently verified. The company asserts that Beam achieves results comparable to Z.ai's GLM-5.2, which has around 744 billion parameters with 40 billion active, and would surpass current leading open Western models while requiring 3 to 4 times fewer resources for inference. Reflection AI also notes that its own internal evaluations place Beam ahead of Inkling, the open model from Thinking Machines Lab launched in July, across four commonly used coding benchmarks. Inkling is multimodal, while Beam is solely textual. Reflection AI positions itself against closed labs like Anthropic and OpenAI, as well as Mistral, Meta, and Cohere.

Target: Enterprises and Sovereigns, with "AI Factories" in Preparation

Reflection AI targets enterprises and sovereign nations with Beam and its future models. The stated goal is to offer "AI factories," allowing institutions to train Reflection AI's models on their own data to create local and customized systems. The company has begun testing this concept with the Shinsegae group in South Korea. Hedge funds and trading firms are among the interested parties in this type of infrastructure. This approach aligns with the vision advocated by Jensen Huang around "AI factories" and an open AI ecosystem.

Release Timeline and Context of the Announcement

Reflection AI intends to make Beam's weights and all technical specifications public this month, with distribution planned through hyperscalers, neo-clouds, and integrations into open-source libraries upon release. Additional information was shared in a blog post published on Monday, following media reports of an imminent launch over the previous weekend. Reflection AI is based in Brooklyn, was founded in 2024 by two former researchers from Google DeepMind, and is described as being two years old. The company did not respond in time to TechCrunch's requests for information. This announcement could accelerate the race for a Western response to the offerings from DeepSeek, Qwen, and Z.ai.