Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost
Nvidia-backed Reflection AI unveils Beam, its first open-weight model, which it says rivals GLM-5.2 on reasoning with far less inference compute. Weights are due this month.
Reflection AI has unveiled Beam, its inaugural frontier open-weight AI model, aiming to challenge leading Chinese models while offering a more cost-effective solution. The two-year-old company asserts that Beam matches the performance of prominent Chinese open models on advanced reasoning benchmarks, with a fraction of the compute cost and inference time compared to rivals.
This announcement follows Axios' weekend report on the startup's near-launch status, and the company elaborated on Beam's features in a detailed blog post on Monday. Beam, a text-only mixture-of-experts model, is trained on high-compute reinforcement learning for reasoning, coding, and agentic tasks, boasting 501 billion parameters, of which 23 billion are active.
Pre-training involved 23.8 trillion tokens, with a 1 million token context window. While Z.ai’s GLM 5.2 has approximately 744 billion total parameters, with 40 billion active, Reflection's performance claims remain unverified independently. Nevertheless, the company claims Beam scores equally on advanced reasoning benchmarks as Z.ai’s GLM-5.2 and surpasses current leading Western open models while utilizing "3-4x less inference compute."
Positioned against closed research labs like Anthropic and OpenAI, popular open models from Chinese developers, and Western entities such as Mistral, Meta, and Cohere, Reflection intends to target enterprises, the public sector, and developers. Its most significant U.S. competitor may be Inkling, an open model from Mira Murati’s Thinking Machines Lab, released in July.
Beam outperforms Inkling on coding tests in Reflection's benchmarks, although Inkling is multimodal, while Beam is text-only. Founded in 2024 by former Google DeepMind researchers and backed by approximately $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, the company valued itself at a $25 billion pre-money valuation.
Reflection has also secured compute access through deals worth over $7 billion with SpaceX and Nebius for Nvidia's GB300 chips until 2029. The company aims to utilize Beam and subsequent models for enterprises and sovereign nations, proposing "AI factories" that would enable institutions to construct their customized local AI systems by training Reflection’s models using proprietary data.
Nvidia CEO Jensen Huang, a major backer of Reflection, has championed the "AI factory" concept, emphasizing the importance of an open AI ecosystem that would also benefit Nvidia's GPU business.
Written by urgent.news from TechCrunch's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
