Reflection releases Beam, a new open-weight AI model targeting Chinese competitors with cheaper inference

Reflection AI has unveiled Beam, its inaugural open-weight frontier model designed to challenge the performance of leading Chinese AI systems without the hefty computational demands. The two-year-old startup asserts that Beam achieves comparable results to models like DeepSeek and Qwen on complex reasoning benchmarks, but at substantially lower operating costs — a positioning that could intensify competition to develop a Western alternative to Asia's dominant open models.

Model specifications and performance claims#

Beam operates as a text-focused mixture-of-experts architecture, containing 501 billion total parameters with 23 billion active during inference. The model was trained on roughly 23.8 trillion tokens and supports a 1 million token context window. For context, this compares to Z.ai's GLM-5.2, which carries approximately 744 billion total parameters with 40 billion active.

According to Reflection's published benchmarks (which remain independently unverified), Beam achieves reasoning performance levels matching Z.ai's flagship model while simultaneously using 3-4 times less compute during inference. The company characterizes it as a "workhorse model" suited for enterprises, government agencies, and developer communities.

The competitive landscape#

Reflection finds itself navigating a crowded market spanning multiple fronts. The company faces competition from closed-model companies like Anthropic and OpenAI, Western open-weight developers including Mistral and Meta, Chinese labs producing increasingly capable models, and emerging U.S. alternatives such as Thinking Machines Lab's Inkling, released in July. While Reflection's benchmarks show Beam outpacing Inkling on certain coding evaluations, Inkling offers multimodal capabilities that Beam lacks.

Backing and compute infrastructure#

Reflection, launched in 2024 by former Google DeepMind researchers, has assembled significant financial support totaling approximately $4.7 billion from heavyweight backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners. A recent funding round valued the company at $25 billion on a pre-money basis.

The startup has also moved aggressively to secure GPU capacity — essential for developing competitive frontier models. Over the summer, Reflection signed and collectively exceeding $7 billion in value, granting access to Nvidia's GB300 processors through 2029.