TL;DR
Get tech for your team delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
Reflection AI has introduced Beam, its first frontier open-weight model, and says it matches leading Chinese open models on advanced reasoning benchmarks while using three to four times less inference compute. The claims have not been independently verified; Reflection says it plans to release the model weights and technical details this month.
Reflection AI on Monday introduced Beam, its first frontier open-weight AI model, saying it performs on par with leading Chinese open models on advanced reasoning benchmarks while requiring less inference compute. The two-year-old startup’s performance and cost comparisons have not been independently verified, leaving the release of the model’s weights and technical details as the next test of its claims.
Reflection describes Beam as a text-only mixture-of-experts model built for reasoning, coding and agentic tasks. The company says it used high-compute reinforcement learning to train the system, which has 501 billion total parameters, including 23 billion active parameters. It was pretrained on 23.8 trillion tokens and has a reported one-million-token context window.
In its own benchmark comparisons, Reflection says Beam scores on par with Z.ai’s GLM-5.2 on advanced reasoning tests and outperforms leading Western open models, while using “3-4x less inference compute.” The company has also reported that Beam scores higher than Inkling, an open model from Thinking Machines Lab, on four coding tests where both systems have results. That comparison has a limitation: Inkling is multimodal, while Beam handles text only.
Reflection calls Beam a “workhorse model” for enterprises, public-sector institutions and developers. It is positioning the system against Chinese open-model developers, Western open-model providers including Meta, Mistral and Cohere, and closed-model companies such as Anthropic and OpenAI. The company says it will release Beam’s weights and full technical details this month, with access through hyperscalers and neocloud providers and integrations with open-source libraries.
Lower-Cost Inference Is the Pitch
Beam’s announcement adds another well-funded U.S. entrant to competition over open-weight AI models—systems whose weights can be made available for others to run, adapt or integrate. If Reflection’s benchmark and compute claims hold up, the model could give organizations an alternative to both proprietary AI services and open models developed in China, while potentially lowering the cost of running advanced reasoning and coding workloads.
The claim matters particularly to organizations that want to run AI on their own infrastructure or tailor it to internal information. Reflection says its broader goal is to build “AI factories”: customized local systems trained on an institution’s proprietary data. Such deployments could appeal to companies and governments seeking more control over data, model customization and access. But a model’s benchmark results alone do not establish its real-world reliability, total deployment cost or suitability for sensitive workloads.
The release also tests whether a U.S. startup can compete in the open-model market while investing at frontier scale. Reflection is trying to attract users with a combination of claimed performance, lower inference requirements and future local deployment options. Those propositions will be easier to judge after outside researchers and customers can examine the model and reproduce comparisons.
As an affiliate, we earn on qualifying purchases.
Reflection’s Compute and Funding Push
Reflection was founded in 2024 by two former Google DeepMind researchers. According to PitchBook, cited by TechCrunch, the company has raised roughly $4.7 billion from investors including Nvidia, Sequoia Capital and Lightspeed Venture Partners. Its most recent funding round valued it at a $25 billion pre-money valuation.
The company has also pursued access to the computing hardware needed to train and serve large models. This summer, Reflection signed agreements collectively worth more than $7 billion with SpaceX and Nebius to secure access to Nvidia GB300 chips through 2029, according to TechCrunch. The agreements underline the scale of infrastructure behind Reflection’s effort, but do not independently establish Beam’s performance or operating costs.
Reflection has begun testing its sovereign AI factory concept in a partnership with South Korea’s Shinsegae Group. TechCrunch reported that hedge funds and trading firms are among organizations interested in building similar systems. Beam’s initial release could provide a model for those efforts, although details about the partnership’s implementation and results have not been provided in the source material.
““a fraction of the token cost and inference time compute” of rivals”
— Reflection AI, describing Beam in its announcement
As an affiliate, we earn on qualifying purchases.
Benchmark Claims Await Outside Tests
Independent verification is not yet available for Reflection’s comparisons with GLM-5.2 or Western models, or for its claim that Beam uses three to four times less inference compute. The source report does not provide enough detail to establish whether comparisons used identical hardware, settings, prompts or measurement methods. The full technical material and released weights may allow outside researchers to evaluate those questions.
It is also unclear how Beam will perform on practical tasks beyond selected benchmarks, what it will cost to host at different scales, or what licensing terms will govern use of its weights. Reflection did not respond to TechCrunch’s requests for additional information before publication, according to the report. Details about the Shinsegae partnership’s results and the planned AI factory deployments remain limited.
As an affiliate, we earn on qualifying purchases.
Weights and Technical Details Due
Reflection says it plans to publish Beam’s weights and full technical details during October, with distribution through hyperscalers and neoclouds and integrations across open-source libraries at launch. Those releases should give developers and independent evaluators an opportunity to inspect the model and test the company’s benchmark and compute claims.
Further indicators will include whether organizations adopt Beam for production workloads, how it performs against comparable models under transparent testing, and whether Reflection provides concrete information about inference costs and local deployments. Until then, the central performance and efficiency claims remain company-reported rather than independently established.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Reflection AI’s Beam?
Beam is Reflection AI’s first frontier open-weight model. The company describes it as a text-only mixture-of-experts system for reasoning, coding and agentic tasks.
How large is Beam?
Reflection reports 501 billion total parameters, with 23 billion active parameters. It says the model was pretrained on 23.8 trillion tokens and has a one-million-token context window.
Has Beam’s performance been independently confirmed?
No independent verification is included in the announcement material reported by TechCrunch. Reflection’s claims that Beam matches GLM-5.2 on advanced reasoning benchmarks and uses three to four times less inference compute remain company-reported claims.
When will Beam’s weights be available?
Reflection says it will release the weights and full technical details during October 2026, with distribution through hyperscalers and neoclouds. The precise release date was not specified.
What does Reflection mean by an AI factory?
Reflection uses “AI factory” to describe a customized local AI system built for an organization, using Reflection models trained on that institution’s proprietary data. The company has begun testing the concept with South Korea’s Shinsegae Group.
Source: rss
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
