Reflection AI Launches Beam, Open-Weight Model Rivaling Top Chinese AI at Lower Inference Cost

Image: TechCrunch · Source

Brooklyn startup Reflection AI unveils Beam, a 501-billion-parameter open-weight language model claiming comparable reasoning performance to China's GLM-5.2 with 3-4x less compute. Beam targets enterprise and sovereign applications, with weights releasing this month.

Reflection AI, a Brooklyn-based startup founded by ex-Google DeepMind researchers, has officially launched Beam, its first open-weight AI model designed to compete with leading Chinese language models such as Z.ai's GLM-5.2. The company announced Beam in a detailed blog post, following initial reporting from Axios that suggested its imminent debut. Beam is a text-focused mixture-of-experts model trained with extensive reinforcement learning, excelling at reasoning, coding, and autonomous agent tasks while dramatically reducing inference compute costs.

Beam contains 501 billion parameters with 23 billion active parameters during use and was pre-trained on 23.8 trillion tokens. Notably, it features a 1 million token context window, allowing it to consider a vastly extended input compared to many competitors. By comparison, GLM-5.2 has approximately 744 billion total parameters and 40 billion active parameters. Reflection claims Beam matches GLM-5.2's performance on advanced reasoning benchmarks while requiring 3 to 4 times less inference compute, though these claims have yet to be independently verified.

Positioned as a cost-effective "workhorse model," Beam targets enterprises, public sector organizations, and developers seeking open models with robust AI capabilities. Reflection is also competing with Western AI labs such as Anthropic, OpenAI, Mistral, Meta, Cohere, and newer entrants like Mira Murati's Thinking Machines Lab. Reflection's benchmarks indicate Beam outperforms the Thinking Machines' Inkling model on coding tasks, although Inkling is multimodal and Beam is text-only.

Reflection has secured significant funding totaling approximately $4.7 billion from prominent investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, with a recent valuation at $25 billion. The company has also locked in substantial compute resources through deals with SpaceX and Nebius worth over $7 billion, ensuring access to Nvidia's latest GB300 GPUs through 2029. This infrastructure is critical for training large-scale frontier models capable of competing with major closed-source providers.

Looking ahead, Reflection aims to build “AI factories” — platforms enabling enterprises and sovereign nations to create localized, customized AI systems by training on proprietary data. This vision aligns with Nvidia CEO Jensen Huang’s advocacy for open AI ecosystems powered by Nvidia hardware. Early partnerships include testing with South Korea’s Shinsegae Group, and interest reportedly includes hedge funds and trading firms seeking bespoke AI.

Beam’s model weights and full technical documentation are expected to be released within October 2026, with distribution planned through hyperscaler platforms and cloud providers alongside integration into open source libraries. This release could mark a significant development in the competition between Western and Chinese open large language models, especially given Beam’s promising claims of efficiency and performance.