Reflection launches Beam, an open-weight frontier model with 501B parameters and 1M token context

Reflection, a two-year-old Brooklyn-based startup founded by former Google DeepMind researchers, officially launched Beam, its first open-weight frontier model. The company describes Beam as a text-only mixture-of-experts model trained with high-compute reinforcement learning to improve reasoning, coding, and agentic task performance.

Beam has 501 billion total parameters with 23 billion active parameters, a 1 million token context window, and was pretrained on 23.8 trillion tokens. Reflection positions the model as a "workhorse" for enterprises, the public sector, and developers, aiming to rival closed labs like Anthropic and OpenAI as well as open models from Chinese developers such as Z.ai.

Reflection claims Beam matches the performance of leading Chinese open models on advanced reasoning benchmarks while using 3–4x less inference compute, though these performance claims have not been independently verified. The company also asserts that Beam outperforms today’s leading Western open models on its own benchmarks, but it did not provide third-party validation.

The startup has raised roughly $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, and secured multi-billion-dollar compute agreements with SpaceX and Nebius to access Nvidia’s GB300 chips through 2029. Reflection is targeting sovereign nations and enterprises with its "AI factories" concept, which would enable institutions to train Reflection’s models on proprietary data to build customized, local AI systems.

Reflection plans to release Beam’s weights and full technical details this month, with distribution through hyperscalers, neoclouds, and open source libraries at launch. The company did not respond to TechCrunch’s requests for additional information by publication time.