Beam: Reflection AI Takes Aim at China’s Lead in Open A.I. Models

Set Trending Topics as a preferred source on Google.

An American start-up wants to beat Chinese A.I. labs on their home turf: Reflection AI, the Nvidia-backed company founded by two former DeepMind researchers, has just unveiled Beam, its first open-weight model. With 501 billion parameters, Beam is designed to keep pace with China’s best open models in coding and A.I. agents while using far less computing power. There are no independent benchmarks yet, however, because the weights have not been released.

What Beam Can Do

Beam is a so-called mixture-of-experts model: Only 23 billion of its 501 billion parameters are active per token, which makes it cheaper to run. The model is text-only but supports a context window of up to one million tokens. A reasoning effort parameter lets users decide whether Beam should answer briefly and economically or think longer and more thoroughly.

According to Reflection AI, Beam is still undergoing final red-teaming and evaluations. Selected users can get early access through a waitlist. The weights are set to follow later this month under the permissive Apache 2.0 license, together with a technical report, a model card and tools for running and fine-tuning the model.

Massive Computing Power for Training

For pretraining, Reflection used 23.8 trillion tokens from the web, public sources and licensed datasets. The base run on 6,144 Nvidia GB300 GPUs took less than four weeks. It was followed by a reinforcement learning run of the same length on 10,500 GB300 chips, during which the model completed more than 100 million rollouts across roughly one million training environments. Reflection calls it one of the largest reinforcement learning runs conducted by any open lab to date and says capabilities were still improving when training ended.

Efficiency is the main selling point. On demanding reasoning tasks, Beam is said to match GLM-5.2 from Z.ai while needing three to four times less compute at inference. Compared with models of more than two trillion parameters, such as Alibaba’s Qwen 3.8 Max, the advantage is said to be even larger.

Benchmarks So Far Come From Reflection Itself

Reflection describes Beam as competitive with GLM-5.2 and close to Qwen 3.8 Max on coding and agentic tasks. In terms of raw capability, the start-up concedes that top models like Moonshot AI’s Kimi K3 remain ahead. An excerpt from the published scores (NR: not reported):

Beam’s scores were reported by Reflection, while the competitors’ figures come partly from Artificial Analysis and DataCurve. Beam is not yet listed in the Artificial Analysis Intelligence Index, which combines results from numerous tests. On DeepSWE and Terminal Bench, Beam trails not only Kimi K3 and GLM-5.3 but also DeepSeek V4.1 Flash, which scores 39 points in the Artificial Analysis index, well behind the open-weight leaders. On the tests where Beam reports its best results, scores for the strongest Chinese models are missing.

A $25 Billion Valuation Before the First Model

Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou, who had worked on Gemini and AlphaGo at Google DeepMind. Last year, the start-up raised $2 billion (about 1.7 billion euros) at an $8 billion valuation in a round led by Nvidia. According to The Wall Street Journal, it is now valued at around $25 billion (about 21.4 billion euros), with Sequoia and Lightspeed among its backers alongside Nvidia. Reflection rents part of its computing power from SpaceX: The deal for Nvidia servers at the Colossus 2 data center in Memphis is worth around $6.3 billion (about 5.4 billion euros) to Elon Musk’s company through 2029.

Investors like to call Reflection the “DeepSeek of the West.” The start-up is part of Nvidia’s Nemotron Coalition, which aims to establish open models as the foundation for sovereign A.I. in allied countries, and is working with the South Korean conglomerate Shinsegae on Korean-language models. Critics recently pointed out that Reflection reached its high valuation without releasing a single model. With Beam, the start-up now has to prove that the expectations are justified.

Strong Competition From China and the U.S.

Chinese developers currently set the pace among open-weight models. Xiaomi’s MiMo-V2.6-Pro leads the open models on Artificial Analysis, followed by Z.ai’s GLM-5.3, Alibaba’s Qwen 3.8 Max and Moonshot AI’s Kimi K3, which triggered a new “DeepSeek moment” on stock markets over the summer. Competition is growing in the West as well: Mira Murati’s Thinking Machines Lab has released Inkling, an open model with 975 billion parameters, and Nvidia offers its own open model family, Nemotron, which Salesforce, among others, builds on. Both trail Beam in Reflection’s own comparisons. Reflection calls Beam the first model in a series and says its successor is already in training.