NEW YORK — A heavily funded U.S. artificial-intelligence startup backed by Nvidia has launched a new open-weight model designed to challenge China’s growing dominance in customizable AI, turning a technical model release into another front in the escalating U.S.-China technology race.
Reflection AI unveiled Beam, its first open-weight frontier model, on October 5, saying the system can compete with leading Chinese models on coding, reasoning and autonomous-agent tasks.
Beam contains 501 billion total parameters, but uses only about 23 billion active parameters for each query, a sparse architecture designed to deliver strong performance without activating the entire model every time.
Reflection says Beam was trained on 23.8 trillion tokens and later subjected to an enormous reinforcement-learning phase using 10,500 Nvidia GB300 GPUs over four weeks, generating more than 100 million training rollouts.
That makes Beam technically ambitious.
But its strategic importance may be even bigger.
The United States has dominated the world’s most powerful proprietary AI systems through companies such as OpenAI, Anthropic and Google.
China, meanwhile, has built a powerful position in a different part of the market:
open and open-weight AI models that businesses can download, customize and run on their own infrastructure.
Reflection wants to close that gap.
What makes Beam different
Most consumers interact with AI through closed services.
ChatGPT, Claude and many commercial Gemini products operate through systems controlled by the companies that built them.
Users can access the intelligence.
They generally do not receive the model weights themselves.
Open-weight models work differently.
A company can potentially download the model, operate it on its own servers and customize it without sending every request through the original developer.
That is especially attractive to:
governments;
defense organizations;
banks;
healthcare systems;
and large companies handling sensitive information.
Reflection says its entire strategy is built around giving customers greater control over models, software and infrastructure rather than forcing them into permanent dependence on one proprietary AI provider.
China has become extremely strong in this market
This is where the geopolitical competition becomes important.
Chinese companies including Alibaba, Z.ai, Moonshot AI and DeepSeek have built increasingly capable models that developers can customize or deploy more independently than many leading American proprietary systems.
Reuters reported that Beam is specifically intended to compete with Chinese models including Z.ai’s GLM-5.2 and Alibaba-linked model families.
The Financial Times reported that Chinese models have gained significant share in public and customizable AI because they combine strong capabilities with relatively low operating costs.
That has created an awkward situation for Washington.
America may still have many of the world’s most advanced closed AI systems.
But companies around the world that want cheaper, customizable models have increasingly looked to China.
Reflection is trying to give them a U.S. alternative.
Reflection says Beam can match a leading Chinese model
Reflection says Beam performs competitively with Z.ai’s GLM-5.2 on coding and agentic workloads.
Reuters similarly reported that Beam’s performance appears competitive with GLM-5.2 and approaches Alibaba’s Qwen family on some tasks.
But this needs an important qualification.
These benchmark comparisons come largely from Reflection’s own testing.
The company has not yet released every technical detail needed for the broader research community to independently validate all of its performance claims.
Reflection also says Beam is still undergoing final evaluations and red-team testing.
So Beam should not yet be described as definitively beating China’s best open models.
It is more accurate to say Reflection has presented evidence that the model may be competitive with them.
The efficiency claim may matter as much as raw intelligence
One of Reflection’s biggest selling points is cost.
The company says Beam can deliver strong coding and agent performance with relatively efficient inference because only part of its 501-billion-parameter architecture is active during each request.
MarketWatch reported that Beam activates around 23 billion parameters per query, helping reduce computational demand compared with much larger dense models.
That matters because AI economics are increasingly becoming a competition over cost per useful answer, not simply benchmark scores.
A model can be extraordinarily intelligent and still struggle commercially if it costs too much to run.
Chinese developers have been especially aggressive about lowering inference costs.
Reflection is trying to compete on the same battlefield.
Beam was trained with an extraordinary amount of Nvidia hardware
Reflection’s infrastructure also highlights why Nvidia remains central to almost every part of the global AI race.
The company says it pretrained Beam on a cluster of 6,144 Nvidia GB300 NVL72 GPUs, completing the core pretraining phase in less than four weeks.
The reinforcement-learning phase was even larger.
Reflection used 10,500 GB300 GPUs for four weeks and generated more than 100 million rollouts across coding, science and autonomous-agent environments.
The company says it also used around 1.3 billion sandbox environments while training and grading the system.
That gives a sense of how capital-intensive frontier AI development has become.
Building a competitive model now requires not only research talent but enormous computing infrastructure.
Nvidia has a strategic reason to back Reflection
Nvidia is not simply selling chips to the startup.
It is also an investor.
The FT reported that Nvidia has invested roughly $800 million in Reflection as part of its effort to strengthen alternatives to dominant closed labs and help build a competitive U.S. open-model ecosystem.
That makes strategic sense for Nvidia.
The more AI developers exist, the more companies need Nvidia hardware.
A world dominated entirely by a handful of giant labs could eventually reduce the number of independent customers buying large-scale computing infrastructure.
Supporting new frontier-model companies expands Nvidia’s ecosystem.
Reflection therefore serves two purposes:
it competes with China;
and it creates another major buyer and user of Nvidia technology.
Reflection has raised billions despite being only two years old
Reflection was founded in 2024 by former Google DeepMind researchers, including chief executive Misha Laskin.
The company initially focused heavily on autonomous software engineering before expanding into frontier foundation models.
Its valuation has risen extraordinarily quickly.
Market data providers put its latest confirmed financing valuation at roughly $25 billion to $27.5 billion, following a multibillion-dollar funding round earlier this year.
That is a remarkable valuation for such a young company.
It reflects how much strategic value investors now attach to companies capable of training frontier models independently.
Reflection is increasingly tied to Washington
The company is not positioning itself simply as another commercial chatbot developer.
Its strategy specifically targets governments and strategic infrastructure.
Reflection says its technology is designed for public-sector organizations that want to operate models on sovereign infrastructure, maintaining control over data, governance and long-term deployment.
The FT reported that Reflection has relationships with parts of the U.S. government, including defense and energy agencies, and has also pursued agreements with allied countries.
That makes Beam part of a broader U.S. effort to reduce dependence on foreign AI systems for sensitive applications.
The Pentagon is increasingly interested in commercial AI
Defense agencies worldwide are moving rapidly toward AI for:
cybersecurity;
intelligence;
software development;
logistics;
planning;
and autonomous systems.
For those applications, open-weight models can be attractive because the military may not want highly sensitive information passing through an outside commercial API.
A model that can run inside classified or government-controlled infrastructure provides greater operational control.
That is exactly the market Reflection is targeting.
South Korea is becoming another important battleground
Reflection has also been expanding internationally.
Its own news page highlights partnerships and infrastructure agreements in South Korea, including plans linked to large-scale AI computing capacity.
That comes as Seoul itself prepares a huge sovereign AI push.
South Korea announced on October 6 that it plans to invest about 4.7 trillion won, or $3.5 billion, in developing a frontier AI model beginning in 2027.
That shows the AI race is no longer only America versus China.
Countries increasingly want domestic or allied models they can control.
Reflection is betting that U.S.-aligned nations will prefer an American open-weight alternative over Chinese systems.
“Open” does not always mean fully open source
Another important distinction is terminology.
Beam is described as open-weight.
That means users are expected to gain access to the trained parameters that allow the model to operate.
But open-weight AI is not automatically identical to completely open-source software.
True open source can also involve releasing:
training code;
datasets;
software tools;
methodology;
and licensing rights.
Reflection says its broader mission is to go further over time by publishing research and releasing tools that allow developers to customize its models.
For Beam specifically, however, the weights are still undergoing final evaluation before broad release.
Safety becomes more complicated when weights are downloadable
Open models create benefits.
They also create risks.
Once powerful model weights are publicly available, the original developer loses some ability to control how the model is modified or deployed.
Closed AI providers can often impose safety restrictions through centralized servers.
Open-weight systems can potentially have those restrictions altered.
Reflection acknowledges this challenge.
Its safety framework says an open-weight model should not be released if it creates unacceptable new systemic risks compared with models already publicly available.
The company says each release must undergo evaluation of potential misuse and external risk.
That is one reason Beam’s public release has not happened immediately.
China faces similar safety questions
Chinese open-model developers are dealing with the same issue.
Reuters Breakingviews recently noted that Chinese AI companies have historically published less detailed safety information than leading U.S. labs, although companies such as Z.ai have begun increasing their safety disclosures.
That matters because open models can travel globally almost instantly.
A model built in Beijing, Hangzhou, New York or San Francisco can be downloaded and deployed elsewhere.
Traditional export controls therefore become much harder once model weights are freely available.
That is forcing governments to reconsider what national AI security actually means.
Washington faces a strategic contradiction
The United States wants to remain ahead of China in AI.
But it also worries about dangerous models spreading too freely.
Those two goals can conflict.
Keeping powerful U.S. models closed may make them easier to control.
But it can also encourage foreign developers to build on cheaper Chinese alternatives.
Releasing powerful U.S. open models helps American technology spread internationally.
But it reduces centralized control.
Reflection is effectively arguing that America cannot allow China to dominate the open ecosystem simply because open models create safety challenges.
Chinese models gained ground partly because they were available
This may be the most important lesson.
A company does not always choose the technically strongest model.
It may choose the one it can:
download;
modify;
host locally;
fine-tune;
and operate cheaply.
China’s open-model strategy has benefited from that reality.
Developers who do not want permanent dependence on OpenAI, Anthropic or Google have alternatives.
That can gradually create ecosystems around Chinese technology.
Washington increasingly views that as both a commercial and geopolitical problem.
Open AI can create standards that become difficult to displace
Technology history shows why.
Linux became foundational because developers could build around it.
Open-source databases became infrastructure.
Public programming frameworks became industry standards.
Once large developer communities form around a technology, replacing it becomes difficult.
Reflection’s philosophy explicitly compares open AI with technologies such as the internet and Linux, arguing that shared systems often produce faster innovation and larger ecosystems.
China has already recognized that advantage.
Reflection wants U.S. companies to compete more aggressively for the same developer base.
Beam is optimized heavily for coding and AI agents
The model is not being marketed primarily as another consumer chatbot.
Reflection emphasizes:
software engineering;
tool use;
reasoning;
STEM problems;
and autonomous agents.
That reflects where many investors believe the next economic value from AI will come from.
Chatbots attracted consumers.
Agents are supposed to do work.
They can potentially:
write software;
use digital tools;
search databases;
manage workflows;
and complete multistep tasks.
If agentic AI becomes commercially important, coding and reasoning benchmarks may matter more than conversational personality.
But benchmarks can mislead
Every major AI lab now publishes charts showing its latest model beating competitors.
Those results should be treated carefully.
Benchmark performance can depend on:
prompt design;
testing methodology;
model settings;
data contamination;
and which tasks the developer chooses to highlight.
The true test comes when independent researchers and customers deploy the model at scale.
Beam has not yet undergone that full market test.
That makes Reflection’s launch significant—but still preliminary.
Reflection already has enormous infrastructure commitments
The startup is also spending aggressively on compute.
Its company news page points to a $1 billion AI-capacity agreement with Nebius and a computing deal with SpaceX worth up to $6.3 billion.
Those commitments underline how expensive this competition has become.
A startup can raise billions and still need billions more for computing.
That creates another risk.
If model performance fails to translate into revenue, huge infrastructure commitments can become financial burdens.
A $25-billion valuation creates enormous expectations
Reflection is still young.
Yet investors have already assigned it a valuation measured in tens of billions of dollars.
That means Beam does not merely need to be technically impressive.
Reflection must eventually prove it can build a durable business.
Potential revenue sources include:
government contracts;
enterprise AI deployments;
infrastructure services;
developer tools;
and international sovereign-AI projects.
The challenge is converting open technology into profits without losing the openness that makes the product attractive.
The U.S. still dominates closed frontier AI
It is important not to exaggerate America’s weakness.
OpenAI, Anthropic and Google remain among the most powerful AI developers in the world.
American companies dominate much of the advanced semiconductor and cloud infrastructure supporting AI.
Nvidia itself remains central to the global ecosystem.
The competitive gap is narrower and more specific:
China has built exceptional momentum in low-cost, customizable open-weight models.
Reflection is trying to address that particular weakness.
China is unlikely to stand still
Beam may narrow the gap.
Chinese labs will keep releasing new models.
Reuters has reported rapid progress from companies including Z.ai and Alibaba even as they operate under U.S. chip restrictions.
That makes the race unusually fast.
A model that looks competitive today may be surpassed within months.
There may never be a permanent leader.
Instead, the advantage may belong to whoever can repeatedly train capable models at the lowest cost while building the largest developer ecosystem.
The real battle is over who controls the AI stack
This is why Beam matters beyond benchmark tables.
Closed AI concentrates power inside a small number of laboratories.
Open-weight AI distributes more control to:
developers;
corporations;
governments;
and infrastructure providers.
China has increasingly offered that second model to the world.
Reflection wants the United States to offer a credible alternative.
The company is effectively making a strategic argument:
countries should not have to choose between expensive U.S. proprietary AI and customizable Chinese models.
America should compete in both.
Beam is an important step—but not yet proof of victory
Reflection has money.
It has Nvidia hardware.
It has government relationships.
And it now has its first frontier open-weight model.
The company says Beam can compete with some of China’s best models while operating efficiently enough to make deployment economically attractive.
But the hardest tests come next.
Independent developers must verify the benchmarks.
Customers must decide whether the economics work.
Safety researchers must examine the model.
And Reflection must show it can keep pace when Chinese rivals release their next generation.
Beam may be America’s strongest new answer to China’s open-model surge.
But the larger race will not be decided by one launch.
It will be decided by which ecosystem developers, companies and governments around the world ultimately choose to build on.
