Reflection debuts Beam, an open-weight AI mannequin to rival Chinese language fashions at decrease compute value

Reflection AI is formally unveiling Beam, its first frontier, open-weight AI mannequin. The 2-year-old, Brooklyn-based startup claims Beam matches the efficiency of main Chinese language open fashions on superior reasoning benchmarks at dramatically decrease prices, a declare that would warmth up the race to construct a Western reply to DeepSeek, Qwen, and Z.ai.

Reflection’s announcement confirms reporting from Axios over the weekend that the startup was near a launch. The corporate shared new particulars in a prolonged blog post Monday, which described Beam as a text-only mixture-of-experts mannequin educated on high-compute reinforcement studying to be efficient at reasoning, coding, and agentic duties at “a fraction of the token value and inference time compute” of rivals.

Beam is a 501-billion-parameter mannequin with 23 billion lively parameters. It was pretrained on 23.8 trillion tokens and has a 1 million token context window. To match, Z.ai’s GLM-5.2 has roughly 744 billion whole parameters with 40 billion lively. 

Reflection’s efficiency claims haven’t been independently verified, however on superior reasoning benchmarks, the corporate says Beam scores on par with Z.ai’s GLM-5.2 and outperforms at this time’s main Western open fashions whereas utilizing “3-4x much less inference compute.” Reflection calls it a “workhorse mannequin” for enterprises, the general public sector, and builders. 

Reflection is positioning itself in opposition to closed labs like Anthropic and OpenAI, in opposition to widespread open fashions from Chinese language builders, and in opposition to Western gamers like Mistral, Meta, and Cohere. Its most direct U.S. rival may be Inkling, the open mannequin from Mira Murati’s Pondering Machines Lab launched in July. Reflection’s personal benchmarks present that Beam outscores Inkling on 4 coding exams the place each report outcomes, however Inkling is a multimodal mannequin and Beam is text-only.

Reflection was based in 2024 by two former Google DeepMind researchers and has raised roughly $4.7 billion from backers, together with Nvidia, Sequoia Capital, and Lightspeed Enterprise Companions, per PitchBook. Its final spherical valued the corporate at a $25 billion pre-money valuation.

The startup has additionally been locking up compute — a key ingredient wanted to coach frontier fashions able to luring prospects away from Anthropic’s and OpenAI’s closed fashions, in addition to the cheaper open-weight fashions from Chinese language labs. This summer season, Reflection signed offers collectively value greater than $7 billion with SpaceX and Nebius to safe entry to Nvidia’s GB300 chips by means of 2029. 

Reflection is aiming Beam and future fashions at enterprises and sovereign nations. The pitch is to construct “AI factories,” a product that might let establishments construct their very own personalized, native AI system by coaching Reflection’s AI fashions on their very own proprietary information. Nvidia CEO Jensen Huang, whose firm backs Reflection, has lengthy championed the “AI manufacturing facility” concept and pushed to strengthen the open AI ecosystem — a imaginative and prescient that might additionally profit Nvidia, whose GPUs would energy these techniques. 

Axios reported that hedge funds and buying and selling companies are amongst these keen to construct such techniques. Reflection has already begun testing the idea of a sovereign AI manufacturing facility partnership with Shinsegae Group in South Korea.

Reflection says it’s going to launch Beam’s weights and full technical particulars this month, with distribution by means of hyperscalers and neoclouds and integrations throughout open supply libraries at launch. 

Reflection didn’t reply in time to TechCrunch’s requests for extra info.