Reflection AI Unveils Beam to Challenge DeepSeek, Qwen and Z.ai
Reflection AI's 501 billion parameter Beam targets Chinese open models, claiming comparable reasoning with three to four times less inference compute.
Summary
Reflection AI unveiled Beam on October 5, 2026, its first frontier open weight model. The text only mixture of experts system has 501 billion total and 23 billion active parameters, was pretrained on 23.8 trillion tokens, and supports a 1 million token context window. High compute reinforcement learning targets reasoning, coding and agentic tasks. Reflection says Beam matches Z.ai’s GLM 5.2, which has about 744 billion total and 40 billion active parameters, beats leading Western open models, and needs three to four times less inference compute. The benchmarks are not independently verified.
Reflection positions Beam against Anthropic and OpenAI, Chinese offerings from DeepSeek, Qwen and Z.ai, and Western open models from Mistral, Meta and Cohere. Its tests show Beam beating Mira Murati’s Thinking Machines Lab model Inkling, released in July, on four shared coding benchmarks, although Inkling is multimodal. Founded in Brooklyn in 2024 by two former Google DeepMind researchers, Reflection has raised roughly $4.7 billion from investors including Nvidia, Sequoia Capital and Lightspeed Venture Partners; its latest round carried a $25 billion pre money valuation.
This summer, Reflection committed more than $7 billion across SpaceX and Nebius deals for Nvidia GB300 access through 2029. It plans customizable local AI factories for enterprises and sovereign nations, with hedge funds and trading firms interested and South Korea’s Shinsegae Group testing a sovereign partnership. Nvidia CEO Jensen Huang champions the concept, which would increase GPU demand. Reflection plans to publish Beam’s weights and full technical details in October 2026 through hyperscalers, neoclouds and open source library integrations.
Positives
- Three to four times less inference compute could make Beam cheaper to operate than similarly capable open models.
- 501 billion total parameters, 23 billion active parameters and a 1 million token window combine scale with sparse activation.
- Four shared coding benchmarks place Beam ahead of Thinking Machines Lab’s Inkling in Reflection’s testing.
- More than $7 billion in SpaceX and Nebius agreements secures Nvidia GB300 access through 2029.
- October 2026 availability will include model weights, technical details, cloud distribution and open source library integrations.
Risks & concerns
- Reflection’s performance and efficiency claims have not been independently verified.
- Beam is text only, while rival Inkling supports multimodal inputs.
- More than $7 billion in compute agreements creates substantial financial exposure through 2029.
- Anthropic, OpenAI, DeepSeek, Qwen, Z.ai, Mistral, Meta and Cohere give Beam a crowded competitive field.