Reflection unveils Beam to challenge Chinese AI models at lower compute cost

Reflection AI is positioning itself as an American competitor to Chinese labs making open AI models. The startup was founded by two former Google DeepMind researchers and has Nvidia's backing, according to Bloomberg Technology.

Its first model, Beam, is the opening product in that effort. Financial Times World reported that Reflection claims Beam is on par with a leading Chinese competitor, putting the launch directly into the competition over open models.

On October 5, 2026, Reflection unveiled Beam. Bloomberg Technology described it as an open-weight model that the company says rivals leading options in both the US and China.

The technical details published that day describe a mixture-of-experts model with 501 billion total parameters and 23 billion active per token, according to DEV Community. Those are different measures: the overall model size and the portion active for each token.

Beam is built for coding, reasoning and agentic work, The Next Web reported. That gives the launch a broader intended use than a model focused only on answering questions.

The central performance claim pairs reasoning ability with lower operating compute. According to TechCrunch, Reflection says Beam rivals GLM-5.2 on reasoning while using far less inference compute, the computation required to run a model rather than train it.

Beam has been announced, but its weights are due later in October, according to TechCrunch. That distinction matters: the unveiling is not the same milestone as making the weights available.

The reported comparisons remain Reflection's claims. Bloomberg Technology and Financial Times World both frame the competitive performance that way. The supplied coverage does not establish independent confirmation of those claims or provide enough benchmark detail to judge how broadly they apply.

The next concrete milestone is the release of Beam's weights this month, as reported by TechCrunch. That will be the point to watch for access beyond the announcement.

The main open question is whether Beam delivers the claimed reasoning performance at substantially lower inference compute. The coverage also leaves the release terms and detailed performance across coding, reasoning and agentic work unresolved.