Cheap and Powerful Western Rival to Chinese AI: Beam from Reflection AI
· AI · Cem Koyluoglu
Reflection AI officially introduces Beam, the first pioneer and open-weight AI model. The two-year-old Brooklyn-based startup claims that Beam reaches the performance of leading open-weight ...
TL;DR Reflection AI has launched Beam, a 501‑billion parameter open‑weight language model that rivals leading Chinese models while cutting inference costs. Beam offers a 1‑million‑token context window and claims to outperform Western competitors with significantly less compute. The startup plans to release weights and open the model to the community later this month.
Key Highlights • Beam is a 501‑billion parameter model with 23 billion active parameters, trained on 23.8 trillion tokens and a 1 million‑token context window. • It claims to match GLM‑5.2’s advanced reasoning scores while using 3‑4× less inference compute and token cost. • Reflection AI has raised ~$4.7 billion, backed by Nvidia, Sequoia, and Lightspeed, and secured future access to Nvidia’s GB300 chips. • The company aims to enable “AI factories” for enterprises, offering local training on proprietary data. • Beam will be released with full weights and integrated into open‑source libraries via hyper‑scale cloud providers.
Reflection AI Unveils Beam, a New Open‑Weight Language Model
Reflection AI, the Brooklyn‑based startup, has officially announced its latest large‑language model, Beam. The company claims that Beam achieves performance on advanced reasoning benchmarks comparable to leading Chinese open‑weight models—such as DeepSeek, Qwen, and Z.ai—while operating at significantly lower cost. The announcement follows a report from Axios that the firm was close to a launch.
Technical Overview
Beam is described as a high‑computational, reinforcement‑learning‑trained mixture‑of‑experts model that relies solely on text. It contains 501 billion total parameters, of which 23 billion are active. The model was pretrained on 23.8 trillion tokens and offers a 1 million‑token context window. For comparison, Z.ai’s GLM‑5.2 model has roughly 744 billion total parameters and 40 billion active parameters.
The company asserts that Beam matches the GLM‑5.2 score on advanced reasoning tests while using “3‑4 times less inference compute.” It also claims to outperform the current leading Western open models by operating with a very small fraction of token cost and inference compute. Beam’s design is positioned as a “model that will take on workloads” for enterprises, the public sector, and developers.
Competitive Landscape
Reflection AI places itself against closed‑lab competitors such as Anthropic and OpenAI, popular Chinese open models, and Western players like Mistral, Meta, and Cohere. Its closest rival in the U.S. market is the Inkling model released by Mira Murati’s Thinking Machines Lab in July. In four coding tests reported by the startup, Beam outperformed Inkling, although Inkling is multimodal while Beam remains text‑only.
Funding and Partnerships
Founded in 2024 by two former Google DeepMind researchers, Reflection AI has raised approximately $4.7 billion from investors that include Nvidia, Sequoia Capital, and Lightspeed Venture Partners. The most recent funding round set the pre‑money valuation at $25 billion.
The startup has secured long‑term agreements to secure the computing capacity needed to train leading models. In 2029, it will have access to Nvidia’s GB300 chips through deals totaling more than $7 billion with SpaceX, Nebius, and Nvidia.
Vision for “AI Factories”
Reflection’s core promise is to enable institutions to build “AI factories” that allow them to train and customize local systems on their own data. Nvidia CEO Jensen Huang has long advocated for the AI‑factory concept, and the partnership aligns with Nvidia’s goal of strengthening the open AI ecosystem by supplying the GPUs that power these systems.
Axios reports that private equity funds and trading firms are also interested in establishing such factories. Reflection has begun testing the sovereign AI‑factory partnership model with South Korea’s Shinsegae Group.
Future Releases
The company says it will publish Beam’s model weights and full technical details later this month. Upon release, the model will be distributed through hyper‑scale cloud providers and next‑generation cloud platforms, with integrations into open‑source libraries.
Reflection AI’s announcement signals a new entrant in the race for cost‑efficient, high‑performance open‑weight language models, positioning itself as a bridge between Western and Chinese AI research and the broader commercial ecosystem.