Reflection has officially launched Beam, a frontier open-weight AI model positioned to match leading Chinese open models on advanced reasoning benchmarks while running at a fraction of the compute cost. The 501-billion-parameter mixture-of-experts model is built to give enterprises, public sector buyers, and developers a Western open-weight option that competes on reasoning, coding, and agentic tasks without the heavy inference bill of today’s largest closed systems.<\/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 it with high-compute reinforcement learning so it could handle reasoning, coding, and agentic workloads, and the company is framing it as a workhorse model for production use rather than a research demo.<\/p>\n
For comparison, Z.ai’s GLM-5.2 carries roughly 744 billion total parameters with 40 billion active. Reflection’s claim is that Beam reaches parity with that model on advanced reasoning benchmarks while using 3-4x less inference compute, though those numbers have not yet been independently verified.<\/p>\n
Reflection is positioning Beam on three fronts at once: against closed labs like Anthropic and OpenAI, against popular open models from Chinese developers such as Z.ai, Qwen, and DeepSeek, and against Western open-weight players including Mistral, Meta, and Cohere. The closest U.S. competitor is Inkling, the open model from Thinking Machines released in July. Reflection’s own benchmarks show Beam outscoring Inkling on four coding tests where both report results, though Inkling is multimodal and Beam handles text only.<\/p>\n
Reflection says Beam’s weights and full technical details will be released this month. Distribution is planned through hyperscalers and neoclouds, with integrations across open source libraries available at launch, which gives developers a direct path to self-host and inspect the model rather than depending on a closed API.<\/p>\n
Reflection was founded in 2024 and has raised roughly $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, with a last-round pre-money valuation reported at $25 billion. To lock in training capacity, the company signed deals collectively worth more than $7 billion this summer with SpaceX and Nebius to secure access to Nvidia’s GB300 chips through 2029, a major commitment to compute at a time when frontier-scale training runs are the single biggest cost driver in the industry.<\/p>\n
Open-weight releases from Chinese labs have moved faster than many Western analysts expected, and the gap on reasoning benchmarks has narrowed to the point where cost, not raw capability, is the main differentiator. A model that can match those results at 3-4x lower inference compute changes the math for any organization weighing a self-hosted open stack against renting access to a closed frontier API. Beam lands in that window.<\/p>\n
Reflection is steering Beam and future models toward a product concept it calls AI factories, which would let enterprises and sovereign nations build customized local AI systems by training Reflection’s models on their own proprietary data. The company has already begun testing the sovereign AI factory idea with Shinsegae Group in South Korea, and reporting indicates hedge funds and trading firms are among the early groups eager to build systems on top of this kind of offering. The pitch lines up with a broader push by Nvidia, which backs Reflection, to grow the open AI ecosystem that runs on its GPUs.<\/p>\n
Independent benchmark runs will be the first real test of Beam’s claims, especially on reasoning and coding tests where open models have historically lagged closed frontier systems. The weight release, the hyperscaler rollout, and the first sovereign AI factory deployment will together show whether Reflection’s compute and capital advantage translates into durable market share against both Chinese open models and the established closed labs.<\/p>\n
Beam is a 501-billion-parameter text-only mixture-of-experts AI model from Reflection with 23 billion active parameters, pretrained on 23.8 trillion tokens with a 1 million token context window and trained with high-compute reinforcement learning for reasoning, coding, and agentic tasks.<\/p>\n
Reflection says Beam scores on par with Z.ai’s GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute, though these performance claims have not been independently verified.<\/p>\n
Reflection says Beam’s weights and full technical details will be released this month, with distribution through hyperscalers and neoclouds and integrations across open source libraries at launch.<\/p>\n