Reflection Beam: An Efficiency-Focused Open Model

Reflection launches Beam, an open-weight language model focused on coding and reasoning. The company claims performance close to Chinese benchmarks with significantly reduced inference consumption and is preparing to release the weights this month.

Reactions Positioning Beam Against Chinese Models

The launch of Beam is presented as an American response to the lead taken by open Chinese models. Beam enters a landscape where open offerings from Meta, Mistral, Nvidia, and Thinking Machines Lab already exist. The competition includes systems published in China by DeepSeek, Alibaba, Z.ai, and Moonshot AI. In March, the Wall Street Journal reported that some investors referred to Reflection as the "DeepSeek of the West," a phrase echoed in many posts on X during the launch. On Monday, the Nvidia AI account on X congratulated Reflection on this launch, and investors, including Deedy Das and Shaun Maguire, expressed their enthusiasm on the same platform.

Partially Activated Design and Claimed Efficiency Gains

Reflection describes Beam as a text model with a total of 501 billion parameters, of which only 23 billion are activated at any given time, a choice aimed at reducing costs and speeding up usage. The company highlights capabilities in coding, multi-step reasoning, and agent tasks. It claims that the model is competitive with Z.ai's GLM-5.2 for coding and agents and approaches Alibaba's Qwen 3.8-Max, while noting that Moonshot AI's Kimi K3 remains ahead in raw capacity. Reflection also states that Beam achieves performance similar to GLM-5.2 on advanced reasoning tests while requiring three to four times less computational resources for inference. Beam is described as a high-performing Western alternative to open-weight models.

Restricted Access Today, Weight Release Announced This Month

Beam is currently in a final security verification phase, and access is limited to a waiting list. Reflection announces its intention to make the model weights public, along with a technical report and tools for developers by the end of the month. The open-weight option allows developers to retrieve the parameters, run them on their own machines, and adapt them, unlike models from OpenAI or Anthropic, which are typically accessible only via an API.

A Well-Funded Young Company with a Vision for Openness

Reflection was founded in 2024 by Misha Laskin and Ioannis Antonoglou, two former researchers from DeepMind. Misha Laskin, CEO, defended an open approach on Monday, stating that intelligence can only be owned if it is open, allowing execution, control, and customization on one's own hardware. The startup claims to have raised billions of dollars, with backing from Nvidia, Sequoia, and Citigroup. In June, it confirmed a funding round valuing the company at $25 billion pre-money. Beam is Reflection's first open-weight model.