Reflection has launched Beam, a 501-billion-parameter open-weight AI model designed to match leading Chinese open models on reasoning, coding, and agentic tasks while using a fraction of the compute. The release positions the two-year-old Brooklyn startup as a Western contender against models such as Z.ai’s GLM-5.2 and DeepSeek, and against closed labs like Anthropic and OpenAI.<\/p>\n
Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters. It was pretrained on 23.8 trillion tokens and supports a 1 million token context window. Reflection trained the model with high-compute reinforcement learning aimed at reasoning, coding, and agentic workloads, and the company says it delivers results at a fraction of the token cost and inference time compute of rivals.<\/p>\n
For comparison, Z.ai’s GLM-5.2 has roughly 744 billion total parameters with 40 billion active. Reflection’s internal benchmarks claim Beam scores on par with GLM-5.2 on advanced reasoning benchmarks and outperforms current Western open models while using 3-4x less inference compute. These performance figures have not been independently verified.<\/p>\n
Reflection is placing Beam in three lanes at once: against closed labs such as Anthropic and OpenAI, against Chinese open models from developers like Z.ai and DeepSeek, and against Western open efforts from Mistral, Meta, and Cohere. The closest U.S. comparison point is Inkling, the open model released in July by Thinking Machines Lab. Reflection’s own benchmarks show Beam outscoring Inkling on four coding tests where both report results, though Inkling is multimodal and Beam is text-only.<\/p>\n
Reflection calls Beam a workhorse model for enterprises, the public sector, and developers. The wider strategy is to package the model inside what the company calls AI factories: customized, locally hosted AI systems that institutions train on their own proprietary data. Reflection has already begun testing the sovereign AI factory concept with Shinsegae Group in South Korea, and reporting indicates hedge funds and trading firms are among those interested in similar setups.<\/p>\n
Frontier training requires locked-in compute, and Reflection has spent the summer securing it. The company signed deals collectively worth more than $7 billion with SpaceX and Nebius to guarantee access to Nvidia GB300 chips through 2029. Nvidia is also a backer, alongside Sequoia Capital and Lightspeed Venture Partners. Reflection has raised roughly $4.7 billion in total, with a last reported pre-money valuation of $25 billion.<\/p>\n
Reflection plans to release Beam’s weights and full technical details this month, with distribution through hyperscalers and neoclouds, and integrations across open source libraries at launch. Because the model is open-weight, the inference cost claim is the central economic pitch: if Beam truly runs at 3-4x less compute than peers on the same tasks, the per-token economics for enterprise deployments could look very different from what closed frontier models offer today.<\/p>\n
Independent benchmark runs will be the first real test of the parity claim with GLM-5.2 and the 3-4x inference compute figure. The Inkling comparison is also incomplete, since Inkling is multimodal and Beam is text-only; a like-for-like coding test, a like-for-like reasoning test, and a published cost-per-token number would settle the marketing in a way a benchmark table cannot.<\/p>\n
Beam is an open-weight, text-only mixture-of-experts AI model from Reflection with 501 billion total parameters and 23 billion active parameters, trained on 23.8 trillion tokens with a 1 million token context window.<\/p>\n
Reflection says Beam matches Z.ai’s GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute. GLM-5.2 has roughly 744 billion total parameters with 40 billion active. The claims have not been independently verified.<\/p>\n
Reflection says it will release Beam’s weights and full technical details this month, distributed through hyperscalers and neoclouds with integrations across open source libraries at launch.<\/p>\n