Reflection AI Introduces Beam, Its First Frontier Open‑Weight Model

Brooklyn‑based startup Reflection AI has officially unveiled Beam, its inaugural frontier‑class open‑weight AI model. The two‑year‑old company, founded by former Google DeepMind researchers, claims that Beam matches the performance of leading Chinese open models on advanced reasoning benchmarks while consuming dramatically less compute. The announcement follows earlier reporting from Axios that indicated the startup was nearing a launch, and Reflection shared additional details in a detailed blog post published on Monday.

Technical Specifications and Architecture

Beam is a text‑only mixture‑of‑experts (MoE) model that boasts a staggering 501 billion total parameters, with 23 billion active parameters during inference. It was pre‑trained on an enormous corpus of 23.8 trillion tokens and features a 1 million‑token context window, allowing it to handle very long documents or codebases in a single pass. For perspective, Z.ai’s GLM‑5.2 model, a recent Chinese open model, has roughly 744 billion total parameters and about 40 billion active parameters.

The model’s MoE design enables it to activate only a subset of its parameters for each token, which Reflection says translates into a “fraction of the token cost and inference time compute” compared with dense rivals. The company emphasizes that Beam was trained using high‑compute reinforcement learning to sharpen its abilities in reasoning, coding, and agentic tasks.

Performance Claims and Competitive Positioning

Although independent verification of Reflection’s benchmarks is pending, the company asserts that Beam scores on par with Z.ai’s GLM‑5.2 on advanced reasoning benchmarks and outperforms today’s leading Western open models while using only 3‑4 times less inference compute. In direct comparisons with Inkling—the open model released in July by Mira Murati’s Thinking Machines Lab—Reflection’s own benchmarks show Beam outperforming Inkling on four coding tests where both models report results. Inkling is multimodal, whereas Beam is text‑only, which Reflection notes as a trade‑off.

Reflection positions Beam as a “workhorse model” aimed at enterprises, public‑sector agencies, and developers who need strong reasoning capabilities without the prohibitive compute costs associated with closed‑source offerings from Anthropic, OpenAI, or the larger Chinese models such as DeepSeek, Qwen, and Z.ai.

Funding, Compute Partnerships, and Valuation

Reflection was founded in 2024 and has raised approximately $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, according to PitchBook. The most recent funding round valued the company at a $25 billion pre‑money valuation. To support the training of frontier models, Reflection has secured massive compute agreements: this summer it signed deals worth more than $7 billion collectively with SpaceX and Nebius to guarantee access to Nvidia’s GB300 chips through 2029.

These partnerships are intended to lock in the hardware needed to train models capable of luring customers away from both the premium closed models of Anthropic and OpenAI and the cheaper open‑weight alternatives emerging from Chinese labs.

Vision of AI Factories and Enterprise Focus

Reflection is targeting Beam and its forthcoming models at enterprises and sovereign nations, promoting the concept of “AI factories.” This product would allow institutions to build customized, local AI systems by training Reflection’s models on their own proprietary data. Nvidia CEO Jensen Huang—a backer of Reflection—has long championed the AI factory idea, arguing that it strengthens the open AI ecosystem while driving demand for Nvidia’s GPUs.

Axios reported that hedge funds and trading firms are among those eager to adopt such sovereign AI factories. Reflection has already begun testing the concept with Shinsegae Group in South Korea. The startup plans to release Beam’s weights and full technical details this month, distributing them via hyperscalers, neoclouds, and integrations across popular open‑source libraries.

Reflection did not respond to a request for comment from TechCrunch in time for publication.