Reflection AI unveils Beam, an open-weight model that rivals Chinese labs on less hardware

Reflection AI, a two-year-old Brooklyn startup, has officially launched Beam, a text-only mixture-of-experts model with 501 billion total parameters — 23 billion active per forward pass. The model was pre-trained on 23.8 trillion tokens and ships with a one-million-token context window. For comparison, Z.ai's GLM-5.2 carries roughly 744 billion total parameters with 40 billion active. Reflection describes Beam as a "workhorse model" for enterprises, the public sector and developers, and says it was trained with high-compute reinforcement learning to excel at reasoning, coding and autonomous-agent tasks.

According to the company's internal benchmarks, Beam matches GLM-5.2 on advanced reasoning tests and outperforms the current leading Western open models while using "3-4x less inference compute." Those claims have not been independently verified, and Reflection has not published full performance figures against standard baselines. The company positions itself against closed labs such as Anthropic and OpenAI, Chinese open models, and Western open-weight players including Mistral, Meta and Cohere.

The most direct comparison is with Inkling, the open-weight model from Mira Murati's Thinking Machines Lab, released in July. Reflection's benchmarks show Beam ahead of Inkling on four coding tests where both report results, though Inkling is multimodal while Beam is text-only — a fundamental difference in purpose. Reflection also distinguishes itself as open-weight rather than open-source: the weights are released, but the training code and data are not necessarily.

The company was founded in 2024 by two former Google DeepMind researchers and has raised roughly $4.7 billion from investors including Nvidia, Sequoia Capital and Lightspeed Venture Partners, per PitchBook. The latest round valued Reflection at $25 billion pre-money. Over the summer it signed compute deals worth more than $7 billion combined with SpaceX and Nebius to secure access to Nvidia GB300 chips through 2029, a critical ingredient for training frontier models that can pull customers away from closed alternatives and cheaper Chinese options.

The long-term play is selling "AI factories" — a product that would let institutions build sovereign AI systems by training Reflection's models on their proprietary data. Nvidia CEO Jensen Huang, whose company backs Reflection, has long championed the "AI factory" concept and pushed to strengthen the open ecosystem, a move that also serves Nvidia's GPU business. Hedge funds and trading firms are among the early interested parties, according to Axios. Reflection has already begun a pilot "sovereign AI factory" partnership with South Korea's Shinsegae Group.

Reflection plans to release Beam's weights and full technical details this month, distributed through hyperscalers and neoclouds with open-source library integrations available at launch. The company did not respond to TechCrunch's requests for comment in time for publication.