Reflection launches Beam, a 501B open-weight model to rival Chinese labs
What's the deal? Reflection has unveiled Beam, its first open-weight AI model, a 501-billion-parameter sparse mixture-of-experts system with 23 billion active parameters. The Brooklyn-based startup says Beam is built for coding, reasoning, and agentic tasks, and matches leading Chinese open models on advanced reasoning benchmarks at far lower cost.
The details: Beam was pretrained on 23.8 trillion tokens and has a 1 million token context window. Its high-compute reinforcement learning run generated over 100 million rollouts on 10,500 Nvidia GB300 GPUs over four weeks of training.
How it stacks up: Reflection says Beam scores on par with Z.ai's GLM-5.2 on advanced reasoning while using "3-4x less inference compute." GLM-5.2 carries roughly 744 billion total parameters with 40 billion active. Where frontier open models like Kimi K3 lead on raw capability, Reflection pitches Beam's edge as efficiency at inference time.
Why now? The model confirms weekend reporting from Axios that a launch was near. Reflection says it will release Beam's weights, technical report, and developer artifacts later this month, with distribution through hyperscalers and neoclouds.
What's the endgame? Reflection is targeting enterprises and sovereign nations with "AI factories" — a product that lets institutions train its models on their own proprietary data to build customised, local systems. It has already begun testing a sovereign AI factory partnership with Shinsegae GroupDealroom has a profile for this one. Try Dealroom → in South Korea.
Founded in 2024 by two former Google DeepMind researchers, Reflection has raised roughly $4.7 billion from backers including Nvidia, Sequoia CapitalDealroom has a profile for this one. Try Dealroom →, and Lightspeed Venture Partners, per PitchBook. Its last round valued the company at a $25 billion pre-money valuation. This summer it signed deals worth more than $7 billion with SpaceX and Nebius to secure Nvidia GB300 chips through 2029.
What could go wrong? Reflection's performance claims have not been independently verified, and Beam is still undergoing final red-teaming. Its most direct US rival, Inkling from Mira Murati's Thinking Machines Lab, is multimodal, while Beam is text-only.
The signal: Beam is a bid to build a Western answer to DeepSeek, Qwen, and Z.ai, positioning Reflection against both closed labs like Anthropic and OpenAI and open players such as Mistral, Meta, and Cohere. Backing Nvidia CEO Jensen Huang's "AI factory" vision, it also strengthens an open ecosystem that runs on Nvidia's chips.
Read more: TechCrunch, reflection.ai
Image credit: Ch'enMeng
