Key Takeaways
- Chinese open-weight models now compete closely enough to pressure closed-model pricing.
- U.S. labs retain advantages in capital, chips, research depth, and product distribution.
- The divide reflects business incentives, safety choices, and release strategy more than ability.
Chinese Open-Weight Models Have Crossed the Competitive Threshold
Moonshot AI announced Kimi K3 on July 16, 2026, describing it as a 2.8-trillion-parameter mixture-of-experts model with native visual understanding and a one-million-token context window. The model became available through Moonshot’s application programming interface, with the release of downloadable weights scheduled for July 27, 2026. That distinction matters because, as of July 25, Kimi K3 is a commercially accessible Chinese frontier model with an announced open-weight release rather than a model whose full weights are already broadly downloadable.
Kimi K3 follows a sustained sequence of releases from Moonshot AI, Alibaba’s Qwen team, DeepSeek, Z.ai, and other Chinese developers. Those model families now offer strong reasoning, coding, multilingual, multimodal, and agent-oriented capabilities. Their progress has increased concern that U.S. laboratories may retain technical leadership without controlling the models that developers adopt as building blocks.
The 2026 AI Index from Stanford University found that the performance difference between leading U.S. and Chinese models had narrowed to 2.7% as of March 2026. Models from the two countries had traded the lead several times since early 2025. Stanford also found a 3.3% difference between the leading closed model and the leading open model, meaning open models had not displaced closed systems at the frontier but had moved close enough for cost, control, and deployment flexibility to affect purchasing decisions.
That finding requires careful interpretation. Artificial intelligence models do not need to lead every public evaluation to disrupt established businesses. An organization may select a model that performs slightly below a premium closed system if it can operate inside a private cloud, remain under local control, support internal fine-tuning, and deliver predictable costs.
Chinese open-weight models have reached that commercial threshold across a growing set of uses. Their influence comes from a combination of capability and accessibility rather than from an uncontested claim to technical supremacy.
New Space Economy’s examination of American perceptions of China’s AI position describes a related divide. The United States retains measurable advantages in private capital, advanced chip design, cloud platforms, and top-tier model production. China leads or competes strongly in research volume, patent output, industrial deployment, robotics, and the distribution of lower-cost open-weight systems.
That mixed record explains why claims that one country has definitively won the artificial intelligence race remain misleading. The more defensible conclusion is that Chinese companies have reduced the capability gap enough to make release strategy and commercial structure central parts of the contest.
Open Weights Change the Economics of AI Access
Model weights are the learned numerical parameters that determine how an artificial intelligence system processes inputs and generates outputs. When developers release the weights, another organization can download the model, operate it on its own infrastructure, and adapt it for defined tasks.
Open weights do not automatically make a system fully open source. The Open Source AI Definition requires access to the preferred form for modifying an artificial intelligence system, including relevant code and information about the data used to produce the model. Many open-weight releases provide trained parameters and inference code without supplying enough material to reproduce the original training process.
Closed-model providers keep their strongest weights on company-controlled servers. Customers obtain access through an application programming interface, subscription, or managed enterprise service. The provider retains authority over updates, safety filters, account access, capacity, and pricing.
That arrangement can be attractive to customers that want a managed product. The provider handles model serving, hardware maintenance, cybersecurity, scaling, monitoring, and upgrades. Customers do not need to operate large computing clusters or maintain specialist infrastructure teams.
Open weights reverse part of that relationship. The customer gains greater control but accepts more operational responsibility. A downloadable model can reduce recurring application programming interface charges for steady, high-volume workloads, yet the customer still pays for processors, electricity, cloud capacity, engineering, security testing, monitoring, and maintenance.
The economic advantage depends on how the model will be used. A small organization with irregular demand may spend less by purchasing access to a closed service. A large enterprise processing millions of recurring requests may reduce costs by hosting a suitable model internally. A government agency may accept higher operating costs in exchange for greater control over sensitive data.
Open weights also strengthen the customer’s negotiating position. An organization can compare hosting providers, move a workload between compatible models, or reserve expensive closed systems for tasks that require their strongest capabilities. Less demanding work can be routed to locally operated or lower-cost open models.
New Space Economy’s analysis of whether Chinese open-weight models are undermining American companies identifies this shift in bargaining power. As capable models become easier to obtain, economic value can migrate from basic model access toward cloud hosting, processors, cybersecurity, integration, data preparation, workflow software, and industry applications.
The U.S. National Telecommunications and Information Administration examined these tradeoffs in its open-model-weights report. The agency identified benefits for competition, research, customization, and access by smaller organizations. It also examined public-safety, security, and geopolitical risks arising from the unrestricted distribution of increasingly capable systems.
The policy issue is no longer whether open-weight models can affect competition. The stronger question concerns how much commercial value remains with a model developer once customers can download credible substitutes.
Why U.S. Frontier Labs Keep Their Best Systems Closed
American laboratories could release more capable open-weight systems. Their reluctance reflects business incentives, governance choices, and security concerns rather than an absence of technical ability.
Training frontier models requires large computing clusters, advanced processors, specialized researchers, data preparation, evaluation, and extensive post-training. A closed service gives the developer a direct mechanism for recovering part of that investment through subscriptions, application programming interface usage, enterprise contracts, and commercial partnerships.
Controlled delivery also lets providers change system instructions, improve safety filters, patch defects, investigate misuse, adjust capacity, and retire outdated versions. A company can suspend access to a hosted service. It cannot reliably recover every copy of weights that users have downloaded and distributed.
Model-weight security has become an explicit concern for U.S. laboratories. Anthropic’s Responsible Scaling Policy includes controls intended to reduce the risk of unauthorized access and model-weight theft. Its security provisions include restricted permissions, hardware authentication, multiparty authorization, code review, and additional safeguards for higher-capability systems.
Commercial structure reinforces the same choice. A laboratory whose valuation depends heavily on future revenue from model access has limited incentive to release its strongest commercial asset under permissive terms. The calculation changes when the company earns money from advertising, cloud infrastructure, devices, enterprise software, or other products.
Meta can use downloadable Llama models to attract developers, improve its consumer services, strengthen its advertising business, and reduce dependence on rival providers. Google can use Gemma to support its cloud platform, developer tools, mobile operating system, and specialized hardware. A stand-alone model company faces a more direct conflict between open distribution and paid access.
U.S. companies have not abandoned open weights. OpenAI released gpt-oss-120b and gpt-oss-20b on August 5, 2025, under the Apache 2.0 license. OpenAI positioned the models as efficient downloadable reasoning systems suited to local and data-center deployment.
The release showed that a closed-model provider can support open weights without publishing its strongest proprietary system. The same pattern appears across much of the U.S. market: capable open models remain available, but the leading commercial models stay inside controlled services.
Meta remains the most persistent U.S. supporter of broadly distributed general-purpose weights. Google combines downloadable Gemma models with a stronger closed Gemini family. OpenAI provides open reasoning models alongside more capable paid systems. Anthropic keeps Claude weights private and emphasizes controlled access.
The accurate assessment is not that U.S. laboratories cannot build competitive open-weight models. They have chosen not to maintain a broad, frequently updated, frontier-level open-weight lineup matching the release pattern of China’s strongest developers.
That choice protects revenue and governance authority. It also leaves room for Chinese models to become common foundations for international software, research, and enterprise deployment.
China Treats Model Distribution as Industrial Strategy
Chinese companies can benefit from open weights even when direct model revenue remains limited. A widely adopted model can increase demand for domestic cloud services, compatible processors, deployment software, fine-tuning tools, consulting, and technical support.
Alibaba offers a clear example. Qwen belongs to a company with cloud infrastructure, enterprise customers, e-commerce operations, developer platforms, and extensive data-center capacity. Releasing capable Qwen weights can attract users to Alibaba Cloud without requiring each developer to pay Alibaba for basic access to the model.
The Qwen3 family was released in April 2025 with dense and mixture-of-experts models covering different performance and hardware requirements. The family included the 235-billion-parameter Qwen3-235B-A22B model, which activates about 22 billion parameters for each token.
A portfolio of sizes matters because most customers cannot operate the largest model. Smaller models can run on less expensive servers, workstations, or edge devices. Larger models can support demanding reasoning, coding, and agent tasks. By covering several hardware tiers, the developer can build a user community extending beyond major cloud customers.
DeepSeek followed a similar strategy. DeepSeek-V3 uses a mixture-of-experts design with 671 billion total parameters and 37 billion activated for each token. DeepSeek-R1 applied reinforcement learning and post-training methods to strengthen reasoning capabilities while retaining downloadable model weights.
Moonshot AI’s Kimi family extends the pattern into long-context and agent-oriented work. Kimi K2 used one trillion total parameters with 32 billion activated per token. Kimi K3 expands the announced scale to 2.8 trillion parameters and targets long-horizon coding, knowledge work, reasoning, and multimodal tasks.
Open releases can also assist companies operating under computing constraints. External developers create compressed versions, improve inference engines, produce language adaptations, build specialized fine-tunes, and identify software defects. The original developer does not automatically receive every improvement, but the surrounding community can broaden the model’s usefulness.
Distribution also has a geopolitical dimension. New Space Economy’s examination of the AI competition between states describes open-weight models as one part of a contest involving processors, cloud systems, energy, data centers, standards, data, financing, and technical partnerships.
A government that operates a Chinese model on domestic servers avoids continuous dependence on a Chinese application programming interface. It may still adopt model formats, software frameworks, documentation, evaluation practices, and technical tools associated with Chinese providers.
That creates influence without requiring centralized control of every deployment. Chinese models can become embedded in foreign institutions through local hosting, derivative development, and integration with domestic software.
The strategy carries costs. Foreign organizations may reject Chinese models because of procurement restrictions, cybersecurity concerns, training-data questions, licensing terms, political risk, or uncertainty about model behavior. Open distribution can also give foreign competitors access to technical methods and model outputs.
China’s leading developers appear willing to accept those costs in exchange for reach, developer adoption, and a stronger position in the software layer where future applications will be constructed.
Hardware Constraints Reward Efficient Model Design
The open-weight contest is partly a contest over how much capability developers can obtain from available hardware. Access to advanced artificial intelligence accelerators remains uneven because the United States controls many of the highest-performance processor designs and restricts exports of selected products and manufacturing technologies to China.
Those restrictions increase the cost of relying on brute-force scaling. Chinese laboratories have stronger reasons to pursue architectures that reduce active computation, improve training stability, and run across a broader selection of processors.
Mixture-of-experts architecture is one response. A model can store hundreds of billions or trillions of parameters but activate only a subset for each token. Every parameter still requires storage, but the model does not need to use the entire network for each calculation.
DeepSeek-V3 stores 671 billion principal parameters and activates 37 billion for each token. Qwen3-235B-A22B stores 235 billion and activates about 22 billion. Meta’s Llama 4 Maverick stores 400 billion parameters and activates 17 billion.
Other efficiency methods include quantization, improved attention mechanisms, optimized routing, knowledge distillation, smaller task-specific models, and faster inference software. These methods can reduce memory requirements, computation, latency, or serving cost.
Hardware restrictions do not mean China lacks substantial computing capacity. Chinese cloud providers, technology companies, research institutions, and state-supported programs operate large clusters. Domestic accelerator development continues through companies such as Huawei and other semiconductor suppliers.
The issue is comparative access. U.S. laboratories have generally had easier access to the newest Nvidia systems and much larger reported private investment. Stanford estimated that U.S. private artificial intelligence investment reached $285.9 billion during 2025, compared with $12.4 billion in China. Stanford cautioned that private-investment comparisons understate Chinese spending because government guidance funds and public financing use different structures.
A large spending difference can encourage architectural efficiency rather than permanent technical inferiority. Systems designed under tighter computing constraints may become attractive to international users because they cost less to operate after release.
Export controls can still slow access to advanced processors and limit the scale of future training. They may also encourage closer coordination among Chinese processor designers, cloud providers, software developers, research institutions, and model laboratories.
The result is not proof that technology restrictions have no effect. It demonstrates that restrictions can influence the direction of engineering work as well as the pace of development. New Space Economy’s analysis of the AI supply chain explains why processors, memory, fabrication, networking, electricity, cooling, software, and data-center capacity must be assessed together.
U.S. Open Models Remain Strong but Uneven
The United States has several substantial open-weight model families. Their existence makes claims of a total American withdrawal from open development inaccurate.
Meta released Llama 4 Scout and Llama 4 Maverick on April 5, 2025. Both use mixture-of-experts architectures and support text and image inputs. Scout has 109 billion total parameters and 17 billion active parameters. Maverick has 400 billion total parameters and 17 billion active parameters.
Meta trained the released systems with assistance from Llama 4 Behemoth, a much larger teacher model. Meta announced Behemoth as a model still in training and did not release its weights with Scout and Maverick.
Google introduced Gemma 3 in March 2025. The family includes one-billion, four-billion, 12-billion, and 27-billion-parameter versions. Models above the smallest size support visual input, context windows of up to 128,000 tokens, and more than 140 languages.
Google reported in March 2025 that Gemma models had been downloaded more than 100 million times and that developers had created more than 60,000 community variations. Those figures describe adoption at that date and should not be treated as a current July 2026 count.
OpenAI’s gpt-oss family added Apache-licensed reasoning models designed for local and data-center use. The larger model contains 117 billion total parameters and uses a mixture-of-experts architecture. OpenAI also released gpt-oss-safeguard models in October 2025 to support customizable safety classification.
These offerings show that the United States remains competitive in downloadable models. The weakness is inconsistency at the highest capability tier. Meta has the clearest long-term commitment to broadly distributed general-purpose weights. Google supports open models but reserves its strongest Gemini systems for controlled services. OpenAI provides capable open reasoning models beside a stronger proprietary product line. Anthropic does not publicly release Claude weights.
Chinese laboratories have been more consistent in treating open weights as a route to international adoption. American developers generally treat openness as one product tier within a broader commercial strategy.
New Space Economy’s assessment of AI vendor market share shows why model rankings cannot describe the complete market. Nvidia leads high-end accelerators, Amazon Web Services leads global cloud infrastructure share, Microsoft has extensive enterprise distribution, Google controls an integrated platform, and OpenAI has powerful consumer reach. Anthropic, Meta, Oracle, Salesforce, and other companies hold strong positions in defined layers.
China’s open-weight advantage should therefore be described as a distribution and adoption advantage within one part of the artificial intelligence market. It does not erase U.S. strength in processors, cloud platforms, research funding, enterprise software, consumer products, or frontier closed models.
Benchmarks Cannot Measure the Whole Contest
Public benchmarks provide useful evidence, but they cannot settle enterprise procurement or national strategy. Results depend on prompt design, evaluation conditions, tool access, response limits, scoring methods, and the exact model version tested.
The Stanford technical-performance review reported invalid-question rates ranging from 2% to 42% across commonly used evaluations. Stanford also warned that leaderboard position may partly reflect adaptation to the evaluation platform rather than a model’s broad practical ability.
A small score difference should not be treated as a complete measure of economic value. Organizations consider inference cost, latency, uptime, hardware requirements, cybersecurity, auditability, multilingual quality, document handling, coding accuracy, data residency, fine-tuning, and integration with existing software.
A model ranking below the overall leader may perform better on a company’s engineering code, contracts, technical manuals, product catalog, or customer-service language. A smaller model may also deliver lower latency and cost for a repetitive task.
Open weights make private evaluation easier. An organization can test a model against internal material without transmitting that material to an outside provider. It can modify the system for a defined use and measure whether the adapted version improves performance.
Closed providers can offer private environments and contractual safeguards, but customers remain dependent on the provider’s technical interface, service terms, pricing, and release schedule. A provider can replace or retire a hosted model even when a customer prefers the older version.
Developer benchmark claims require caution as well. Companies select evaluations that present their systems favorably. They may test with different prompt formats, reasoning settings, tool access, or computational budgets.
The more defensible interpretation of the 2026 evidence contains two findings. The performance difference between leading U.S. and Chinese systems has narrowed to a small margin. Closed models still retain a modest aggregate advantage over open models.
Both conditions can remain true. Chinese open-weight models do not need to defeat every U.S. closed system to affect the market. They need to become capable enough for coding, document analysis, translation, search, research assistance, customer service, and internal automation.
Once that threshold is crossed, closed providers must justify higher prices through better reliability, stronger tools, easier management, legal assurances, cybersecurity, or superior performance on demanding tasks.
Safety, Security, and Intellectual Property Shape the Divide
An open-weight release cannot be fully reversed. A provider can remove an application programming interface, suspend an account, patch a hosted model, or change its safety rules. It cannot reliably recover every downloaded copy of a model or stop users from altering its behavior.
That difference grows more important as models improve at cybersecurity, scientific reasoning, software development, and autonomous tool use. The policy debate concerns which capabilities can be distributed safely and which safeguards remain effective after the original developer loses control of the weights.
Anthropic has argued for strong protection of advanced model weights and tighter controls on illicit model distillation. Its May 2026 paper on U.S. and Chinese AI leadership attributes part of China’s progress to talent, circumvention of export controls, and extraction of capabilities from U.S. systems. Those claims represent Anthropic’s stated position and remain contested parts of a broader commercial and geopolitical dispute.
Distillation is a common machine-learning method in which one system learns from the outputs of another. It can be used lawfully by a developer training its own model family. Disputes arise when an organization gathers large quantities of outputs from another company’s service in violation of contractual limits or uses them to reproduce protected capabilities.
Open-weight advocates argue that local control and broader access can improve competition, research, and security. Researchers can inspect behavior, build independent safeguards, and operate models inside protected networks. Organizations can test models against their own risk requirements rather than relying solely on a provider’s evaluation.
On July 24, 2026, Nvidia, Microsoft, Meta, IBM, Palantir, Perplexity, and other organizations published an open-weights letter supporting continued U.S. open-model development. The signatories argued that open weights allow startups, universities, public institutions, and businesses to adapt models without paying frontier-model prices for every task.
The U.S. government has also recognized the economic and strategic value of open models. America’s AI Action Plan, released in July 2025, called for leading American open-source and open-weight models. The plan cited benefits for startups, academic research, sensitive government data, and international technical influence.
The unresolved policy problem is how to address measured high-risk capabilities without freezing the market in favor of a small group of closed providers. Rules that are too broad could reduce competition and research. Rules that are too narrow could permit the unrestricted distribution of systems that materially increase misuse risks.
Licensing does not resolve every concern. A permissive license can allow commercial modification without revealing the training data. A restrictive model license may permit research but limit defined commercial uses. Organizations must examine the license, model documentation, evaluation results, security requirements, and applicable law before deployment.
What Enterprises and Governments Should Expect
The market is moving toward a division between controlled frontier services and customizable open-weight infrastructure. Closed systems will remain attractive to customers that want immediate access to the strongest available models, managed security, technical support, service guarantees, and frequent upgrades.
Open models will gain where data control, predictable cost, local operation, customization, and supplier independence carry greater weight. Many organizations will combine both approaches.
A model router can direct difficult reasoning tasks to a premium hosted system and send repetitive work to a smaller local model. Sensitive records can remain inside a private environment, with external services receiving only sanitized information. Organizations can also retain multiple model options to reduce dependence on one provider.
That approach creates new management requirements. Buyers need internal evaluations, access controls, cybersecurity testing, monitoring, model inventories, and procedures for updates. Operating an open model transfers responsibility from the provider to the customer. It does not remove responsibility.
Governments face a similar choice. Sovereign AI refers to the capacity to control artificial intelligence infrastructure, data, models, security, governance, and deployment. Open weights can support that objective because a country can operate a model without continuous access to a foreign application programming interface.
Sovereignty still requires computing capacity, reliable electricity, skilled personnel, cybersecurity, local data, procurement authority, and financial resources. Downloading weights does not create an independent national capability.
Canada offers one middle-power response. The federal government launched AI for All on June 4, 2026. The strategy addresses trust, skills, adoption, domestic infrastructure, Canadian companies, and international partnerships.
New Space Economy’s review of the Canadian AI strategy connects those policy areas to sovereign computing capacity, public procurement, workforce development, safety, and industrial growth.
Countries in Europe, Asia, the Middle East, Africa, and Latin America face related questions. They may want advanced artificial intelligence without placing sensitive public or commercial functions under the continuous control of a foreign company. Chinese open-weight systems can appear attractive because they offer local deployment and lower access costs.
U.S. providers retain strong counteradvantages. They offer established cloud platforms, enterprise support, global developer communities, extensive security programs, and integration with widely used business software. Procurement decisions will depend on trust, cost, technical performance, licensing, political relationships, data rules, and available infrastructure.
For U.S. laboratories, pressure will emerge through pricing and technical influence. A closed model can remain better on difficult evaluations and still lose routine workloads to a cheaper open alternative. A Chinese model can gain international use without generating equivalent direct revenue.
The contest will be determined by performance, cost, distribution, security, hardware access, developer support, licensing, and the ability to turn adoption into sustainable economic value.
Summary
Chinese open-weight artificial intelligence models are closing the gap with U.S. laboratories because Chinese developers have combined improving technical performance with an aggressive distribution strategy. Alibaba, DeepSeek, Moonshot AI, and other companies offer or plan to offer systems that developers can host, adapt, and connect to their own infrastructure.
Near-frontier performance gives those models commercial force even when a closed U.S. system remains stronger on a particular evaluation. Customers may prefer lower cost, data control, customization, or supplier independence over a small benchmark advantage.
The United States retains substantial strengths. It produced more top-tier models in the 2025 data assessed by Stanford, attracted far more reported private investment, hosts many artificial intelligence data centers, and controls leading chip, cloud, and commercial model companies. Meta, Google, and OpenAI also maintain capable downloadable model families.
The weakness lies in release consistency. Most U.S. frontier companies reserve their strongest systems for paid and controlled access. Chinese developers have been more willing to distribute models close to the frontier, using openness to build developer communities and strengthen related cloud, hardware, and software markets.
Closed providers will continue competing through maximum capability, managed service, safety controls, cybersecurity, and convenience. Open-weight developers will compete through cost, customization, local control, and broad distribution.
China’s gains demonstrate that a company can create strategic and commercial influence by distributing the model rather than preserving scarcity around access. U.S. companies can respond, but a stronger open-weight strategy may require them to earn more revenue from hosting, applications, integration, and trusted deployment instead of relying mainly on control of the model interface.
Appendix: Useful Books Available on Amazon
Appendix: Top Questions Answered in This Article
What Is an Open-Weight AI Model?
An open-weight model makes its trained numerical parameters available for download. Users can operate the model on their own infrastructure and may be able to fine-tune it. The release may still exclude training data, development code, evaluation data, and other materials needed to reproduce the system.
Are Open-Weight Models the Same as Open-Source AI?
No. Open weights provide access to learned parameters, but a fully open-source artificial intelligence system requires broader access to the materials needed to study, modify, and share the system. Licenses differ, so users must review commercial-use terms, redistribution rules, and technical documentation.
Have Chinese Models Surpassed U.S. Models?
No single answer applies to every model or task. Stanford found that the leading U.S. model held a 2.7% advantage over the leading Chinese model in March 2026, with the lead changing hands several times since early 2025. Chinese systems may lead on particular languages, tasks, prices, or deployment requirements.
Why Do Chinese Companies Release Strong Models for Download?
Open distribution can attract developers, increase cloud demand, support domestic processors, and spread technical standards. It can also produce external testing, language adaptations, compressed versions, and specialized models. Direct model-access revenue may matter less when the developer earns money from cloud infrastructure or related services.
Why Do OpenAI and Anthropic Keep Their Best Weights Private?
Closed access supports subscription and application programming interface revenue, gives the provider control over updates, and permits monitoring of misuse. It also lets the company patch or retire hosted systems. Downloadable weights would reduce that control and could weaken the scarcity supporting premium pricing.
Does the United States Have Competitive Open Models?
Yes. Meta’s Llama family, Google’s Gemma family, and OpenAI’s gpt-oss models provide capable downloadable options. The difference is that the strongest U.S. commercial models generally remain closed, and American release strategies have been less consistently centered on frontier open weights.
Are Open-Weight Models Free to Operate?
No. The weights may be downloadable without a model-access charge, but users still pay for processors, cloud capacity, electricity, engineering, cybersecurity, testing, and maintenance. Open models become financially attractive when control or recurring workload savings justify those costs.
Do Open Models Offer Better Privacy?
They can. A locally hosted model lets an organization keep prompts and data inside its own environment. Privacy still depends on secure deployment, access controls, logging, software dependencies, cybersecurity practices, and employee behavior.
Why Are Open Weights a National Security Issue?
Open weights can support domestic control and reduce dependence on foreign providers. They can also be modified after release, and the original developer may be unable to prevent misuse. Governments must weigh competition, research, scientific benefit, accessibility, and capability-based security risks.
Will Open Models Destroy Closed-Model Businesses?
That outcome appears unlikely in the near term. Closed providers can retain customers through higher capability, managed service, stronger tools, support, security, and legal assurances. Open models will pressure routine workloads and prices, requiring closed companies to demonstrate the value of controlled access.
Appendix: Glossary of Key Terms
Application Programming Interface
An application programming interface is a software connection that lets one system request functions or data from another. Artificial intelligence companies use these interfaces to provide model access without transferring the weights or the underlying serving infrastructure.
Closed Model
A closed model is an artificial intelligence model whose weights remain controlled by its developer. Users normally access it through an application, subscription, or managed interface, and the provider manages updates, safety systems, capacity, and service conditions.
Distillation
Distillation is a training method in which a smaller or newer model learns from the outputs or behavior of another model. The method is widely used, but disputes arise when outputs are collected in ways that may violate contracts or intellectual property rights.
Fine-Tuning
Fine-tuning is additional training that adapts a general model to a narrower task, language, style, or body of knowledge. It can improve performance for a defined use but may change model behavior and require new safety and reliability evaluations.
Frontier Model
A frontier model is an artificial intelligence system operating near the highest current level of general capability. The label changes as new systems appear and does not mean that the model leads every task, language, cost measure, or safety evaluation.
Inference
Inference is the process of running a trained model to produce an answer, prediction, image, classification, or action. Its cost depends on model size, hardware, token volume, speed requirements, and the efficiency of the serving software.
Mixture of Experts
A mixture-of-experts model stores many specialist parameter groups but activates only a subset for each token. The architecture can provide high total capacity with less computation per request than a dense model containing a similar total number of parameters.
Model Weights
Model weights are the learned numerical parameters that shape how an artificial intelligence system processes input and generates output. Sharing weights lets other organizations operate and adapt a trained model without revealing every part of its original development process.
Open-Source AI
Open-source artificial intelligence provides the materials and permissions needed to use, study, modify, and share a system. The concept extends beyond downloadable weights and can include model architecture, inference code, data information, and other resources needed for meaningful modification.
Sovereign AI
Sovereign artificial intelligence is the capacity of a country or institution to control computing infrastructure, data, models, security, governance, and deployment. Open weights can support sovereignty, but independent capability also requires hardware, energy, talent, financing, and operational expertise.
