Nvidia-backed Reflection AI unveils Beam, open-weight model to rival Chinese models
Version 2: New coverage adds the company’s $25 billion valuation, a reported $6.3 billion deal with SpaceX for Nvidia hardware, training details including a 6,144-GPU cluster, the founders’ names and a waitlist for Beam, with weights now set for release later this month.
In short
Reflection AI unveiled Beam, a 501-billion-parameter open-weight AI model that the New York-based startup says matches leading Chinese open models like Z.ai’s GLM-5.2 at lower compute cost. The Nvidia-backed company says Beam is in final safety testing and available via a waitlist, with weights and technical details to be released later this month. TechCrunch said the performance claims have not been independently verified.
Read the full story 2 min read
Reflection AI unveiled Beam, its first frontier open-weight AI model, the Nvidia-backed, New York-based startup said Monday, according to TechCrunch and Business Insider. The company said Beam is designed for coding, reasoning and agentic tasks, Reuters reported. The launch came after Axios reported over the weekend that Reflection was close to releasing a powerful open-weight system. [ 1 , 2 , 3 , 6 ]
TechCrunch reported Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters, pretrained on 23.8 trillion tokens with a 1 million token context window. SiliconANGLE reported the same 501 billion parameter count and said the model was trained on 23.8 trillion tokens. Reflection says Beam matches leading Chinese open models on advanced reasoning benchmarks at dramatically lower costs, using “3-4x less inference compute,” according to TechCrunch. The company says Beam is on par with Z.ai’s GLM-5.2 and approaches Alibaba’s Qwen 3.8-Max, The Hill and Business Insider reported. TechCrunch said the performance claims have not been independently verified. [ 2 , 4 , 5 , 6 ]
Reflection is positioning Beam against closed labs such as Anthropic and OpenAI, against popular open models from Chinese developers, and against Western players including Mistral, Meta and Cohere, TechCrunch reported. Its most direct U.S. rival might be Inkling, the open model from Mira Murati’s Thinking Machines Lab released in July, according to TechCrunch, which noted Inkling is multimodal while Beam is text-only. SiliconANGLE reported Beam is the first open-source model from a U.S. startup to demonstrate comparable or better performance than advanced Chinese open-source models, though free LLMs still trail frontier models like Anthropic’s Claude Fable 5.1. The Hill reported the release comes amid concerns over China’s build-out of open-source technologies and national security scrutiny, citing Andreessen Horowitz research that 80 percent of developers worldwide using open-source tools build with Chinese models. [ 2 , 4 , 5 ]
SiliconANGLE reported Reflection raised funding at a $25 billion valuation and reportedly signed a $6.3 billion deal with SpaceX to rent Nvidia GB300 NVL72 appliances, using a cluster of 6,144 GPUs to train Beam Base in under four weeks. Business Insider reported the company confirmed a funding round at a $25 billion pre-money valuation in June, with backers including Nvidia, Sequoia and Citigroup. Reflection was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, Business Insider reported. TechCrunch reported the company has raised roughly $4.7 billion from backers including Nvidia and Sequoia Capital. [ 2 , 5 , 6 ]
Beam is in final safety testing and available through a waitlist, Business Insider reported. The Hill and Business Insider reported that the weights and technical details will be released later this month. Reflection CEO Misha Laskin said in a podcast released Monday that “the only way to own intelligence is, by definition, if it's open — because if it's open, you can run it on your stuff, you can own it, you can control it, you can customize it,” Business Insider reported. Axios reported sources expected other open-weight models from Western players this month. [ 1 , 4 , 6 ]
Why it matters
The launch adds a U.S. competitor to an open-model field dominated by Chinese developers, giving companies a cheaper alternative to closed U.S. labs, according to Axios. The Hill reported the release comes amid concerns over China’s open-source build-out and scrutiny of national security implications, while Axios noted open models can be harder to monitor and regulate than frontier models.
Key facts
- Reflection AI unveiled Beam, its first frontier open-weight AI model, on Monday. [ 2 , 3 , 4 , 5 , 6 ]
- Beam has 501 billion total parameters. [ 2 , 5 , 6 ]
- Beam is a text-only mixture-of-experts model that activates 23 billion parameters at a time and was pretrained on 23.8 trillion tokens with a 1 million token context window. [ 2 , 5 , 6 ]
- Reflection says Beam is competitive with Z.ai’s GLM-5.2 and approaches Alibaba’s Qwen 3.8-Max, while using “3-4x less inference compute” than rivals, according to TechCrunch. [ 2 , 4 , 5 , 6 ]
- TechCrunch said Reflection’s performance claims have not been independently verified. [ 2 ]
- Reflection was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, and is backed by investors including Nvidia, Sequoia and Citigroup. [ 2 , 4 , 6 ]
- Beam is in final safety testing and available through a waitlist, and the weights, technical report and developer tools will be released later this month. [ 4 , 6 ]
- The Hill reported, citing Andreessen Horowitz research, that 80 percent of developers worldwide using open-source tools are building with Chinese models. [ 4 ]
Confirmed by several sources
- Reflection AI unveiled Beam, its first open-weight AI model, on Monday. [ 2 , 3 , 4 , 5 , 6 ]
- Reflection is backed by Nvidia. [ 1 , 3 , 4 , 6 ]
- Beam has 501 billion total parameters. [ 2 , 5 , 6 ]
- Reflection says Beam is competitive with Z.ai’s GLM-5.2 and approaches Alibaba’s Qwen 3.8-Max. [ 2 , 4 , 6 ]
- Reflection says Beam uses significantly less compute than rivals, including “3-4x less inference compute.” [ 2 , 5 ]
- The weights and technical details will be released later this month. [ 4 , 6 ]
- Reflection was founded in 2024 by two former Google DeepMind researchers. [ 2 , 4 , 6 ]
- Beam is in final safety testing and available through a waitlist. [ 4 , 6 ]
Still unclear
- Beam’s performance claims have not been independently verified. TechCrunch stated the benchmarks were company claims and were not independently verified.
- Whether Beam initially lags behind the most cutting-edge U.S. models. Axios reported pre-launch expectations that it would lag, while the company claims parity with leading Chinese open models and outperformance of Western open models; the documents do not resolve this.
- Which other Western open-weight models are expected this month, and from which companies. Axios cited sources for the expectation but did not name the models or developers.
- Beam’s exact licensing terms and whether it is fully open-source or open-weight. The documents use “open-source” and “open-weight” and The Hill notes some models may not be open source but can have open weights; no licensing details are given.
- The reported $6.3 billion deal with SpaceX to rent Nvidia GB300 NVL72 appliances. Single-source from SiliconANGLE; other documents do not mention it.
- The timeline for general availability of Beam. The documents say Beam is in final safety testing and available through a waitlist, with weights to be released later this month, but do not give a general availability date.
What local media are saying
Timeline, local time
- 2026-10-04: Axios reports Reflection is preparing to launch an open-weight model, citing sources. [ 1 ]
- 2026-10-05: TechCrunch reports Reflection unveils Beam, citing a company blog post published Monday. [ 2 ]
- 2026-10-05: Reuters reports Nvidia-backed Reflection unveils its first AI model to take on Chinese open models. [ 3 ]
- 2026-10-05: The Hill reports Reflection released Beam, with final evaluations ongoing and weights to be released later this month. [ 4 ]
- 2026-10-05: SiliconANGLE reports Beam details, including 501 billion parameters, a $25 billion valuation and a $6.3 billion SpaceX deal. [ 5 ]
- 2026-10-06: Business Insider publishes an explainer, naming founders and quoting CEO Laskin. [ 6 ]