Reflection AI has officially unveiled Beam, its first frontier open-weight model, claiming it matches leading Chinese models on advanced reasoning benchmarks at a fraction of the inference cost.
The Brooklyn startup, founded in 2024 by two former Google DeepMind researchers, said Monday that Beam is a text-only mixture-of-experts model trained on high-compute reinforcement learning. It is built to handle reasoning, coding, and agentic tasks.
The timing matters. Western labs have spent two years watching DeepSeek, Qwen, and Z.ai set the pace on open-weight performance while keeping inference costs low. Reflection is betting that a Western-built alternative can close that gap without the price tag. By paragraph three, the stakes are clear: if a U.S. startup can match Chinese open models on price and performance, the balance of power in open AI shifts.
The numbers behind Beam
Beam has 501 billion total parameters and 23 billion active parameters. It was pretrained on 23.8 trillion tokens and carries a 1 million token context window. Z.ai's GLM-5.2, by comparison, has roughly 744 billion total parameters with 40 billion active.
Reflection says Beam scores on par with GLM-5.2 on advanced reasoning benchmarks and beats today's leading Western open models while using 3 to 4 times less inference compute. Those claims have not been independently verified.
The company calls Beam a workhorse model for enterprises, the public sector, and developers. It is positioning the model against closed labs such as Anthropic and OpenAI, against Chinese open models, and against Western players such as Mistral, Meta, and Cohere.
Its most direct U.S. rival may be Inkling, the open model from Mira Murati's Thinking Machines Lab released in July. Reflection's own benchmarks show Beam outscores Inkling on four coding tests where both report results. Inkling is multimodal, though, while Beam is text-only. That limitation narrows the audience for developers who need vision or audio in the same stack.
Money and compute
Reflection has raised roughly $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, per PitchBook. Its last round valued the company at $25 billion pre-money.
Compute is the other half of the story. This summer Reflection signed deals collectively worth more than $7 billion with SpaceX and Nebius to secure access to Nvidia's GB300 chips through 2029. Frontier models need that hardware to compete with closed labs and cheaper Chinese open-weight models alike.
The startup is also selling a longer-term vision it calls AI factories. The idea is to let institutions build customized, local AI systems by training Reflection's models on their own proprietary data. Nvidia CEO Jensen Huang, whose company backs Reflection, has long championed the AI factory concept.
Axios reported that hedge funds and trading firms are among those eager to build such systems. Reflection has already begun testing a sovereign AI factory partnership with Shinsegae Group in South Korea. For a company two years old, that is an unusually fast move into enterprise sales.
The factory pitch ties directly to Nvidia's interests. Jensen Huang has argued that every company and country will eventually run its own AI infrastructure, and Reflection's product is aimed at exactly that buyer. If the model is cheap enough to run locally, the pitch gets stronger.
What happens next
Reflection says it will release Beam's weights and full technical details this month, with distribution through hyperscalers and neoclouds and integrations across open source libraries at launch. The company did not respond to TechCrunch's requests for more information.
The launch lands in a market where the price of intelligence keeps falling. Epoch AI measures how fast the cost of a given level of AI performance is dropping across benchmarks covering math, science, and games. If Reflection's 3 to 4 times compute claim holds up, it would be a meaningful data point in that trend.
It is not clear this will hold. Reflection's performance claims are self-reported, and independent evaluation of a 501 billion parameter model takes time and compute that few outside labs have. The text-only limitation also narrows the audience relative to multimodal rivals.
Still, the funding and the compute deals give Reflection room to iterate. Beam is a first model, not a finished product, and the company has signaled more to come. The next question is whether enterprises and sovereign buyers will pay for a Western open-weight alternative when Chinese models already run cheap.
For more on the broader economics of AI compute, see Epoch AI.
The real test comes when the weights land this month and outside researchers can run their own numbers. Until then, Beam is a claim with a large balance sheet behind it. Readers following open models can track more in our AI coverage.
Reflection's bet is simple: cheaper reasoning from a Western lab, sold to institutions that want control over their data. Whether that bet pays off depends on benchmarks nobody outside the company has run yet.
There is also a geopolitical layer. Sovereign buyers in Europe and Asia have been cautious about relying on Chinese open models for sensitive workloads. A Western open-weight alternative with verified performance could find a market even at a slight premium, provided the weights actually ship and the numbers check out.
That caveat matters. Open-weight releases have slipped before, and technical details that look complete in a blog post can take weeks to reproduce. Reflection has promised both weights and full technical details this month, which sets a clear clock.
For now, the most concrete fact is the scale of the backing. Four point seven billion dollars in funding and more than seven billion in compute commitments is a serious entry into a race that Chinese labs have been running with far less capital. The question is not whether Reflection can spend. It is whether the spending produces a model people actually switch to. That answer will not come from a press release. It will come from developers who download the weights, run their own evals, and decide whether Beam earns a place in their stack. The download button is the real launch. now.