Is China Overtaking the US in the AI Race? Open-Source Models vs. Proprietary Giants

Recent weeks have showcased a flurry of advancements, particularly from Chinese AI labs, which have released powerful open-source models like Kimi K2, Qwen3–235B-A22B, Qwen 3 Coder, and ZAI GLM 4.5. In contrast, the US clings to proprietary models, exemplified by OpenAI’s Study Mode in ChatGPT and Anthropic’s Claude Max Pricing.

China’s Open-Source Revolution

China’s AI labs have unleashed a torrent of open-source models, leveraging permissive licenses to empower developers globally. These advancements expose the US’s proprietary approach as restrictive and shortsighted. Let’s examine China’s key releases to contextualize the critique.

Kimi K2: Moonshot AI’s Efficient Giant

Moonshot AI’s Kimi K2, launched in July 2025, is a 1-trillion-parameter Mixture-of-Experts (MoE) model with 32 billion active parameters per forward pass. Its MoE architecture activates only a subset of experts, slashing computational costs while achieving top scores on coding (SWE-Bench, 72%), reasoning (GPQA, 85%), and tool-use (Tau2, 90%) benchmarks. Licensed under a modified MIT, Kimi K2 supports local deployment via vLLM and TensorRT, offering flexibility absent in US models. Its performance rivals Claude Opus 4, exposing the US’s reliance on costly, closed systems.

Qwen3–235B-A22B: Alibaba’s Reasoning Powerhouse

Alibaba’s Qwen3–235B-A22B, released on July 25, 2025, is a 235-billion-parameter MoE model with 22 billion active parameters. Its 262,144-token context length and Thinking mode, which employs chain-of-thought reasoning, yield a 92.3 score on AIME25, surpassing OpenAI’s o4-mini (92.7). On LiveCodeBench v6, it scores 74.1, outpacing Gemini 2.5 Pro (72.5). Licensed under Apache 2.0, it supports efficient inference on consumer hardware, a stark contrast to the US’s cloud-locked models. This accessibility undermines the US’s gatekeeping strategy.

Qwen 3 Coder: Agentic Coding Redefined

Qwen 3 Coder, launched on July 22, 2025, targets autonomous coding with a 35-billion-active-parameter MoE architecture. Its 1-million-token context window handles massive codebases, scoring 70% on SWE-Bench, rivaling Claude Sonnet 4. Integrated with Qwen Code CLI, it executes complex workflows, unlike US models’ API-bound limitations. Its Apache 2.0 license fosters innovation, highlighting the US’s failure to empower developers.

Qwen Small and Medium Models: Scalable Versatility

Alibaba’s Qwen3 family includes dense models from 0.6B to 32B parameters, trained on 36 trillion tokens across 119 languages. The Qwen3–4B matches Qwen2.5–72B-Instruct’s performance, while Qwen3–30B-A3B uses 10% of active parameters for comparable results. These models, deployable via Ollama and LMStudio, democratize AI for resource-constrained users, exposing the US’s elitist, high-cost approach.

StepFun MoE and InternLM Intern S1: Expanding the Ecosystem

StepFun’s MoE model and Shanghai AI Lab’s InternLM Intern S1, both open-weight and permissively licensed, bolster China’s ecosystem. Though less benchmarked, their focus on efficiency and accessibility aligns with China’s strategy to outpace the US’s restrictive models. These releases underscore the US’s missed opportunity to foster global collaboration.

ZAI GLM 4.5 and GLM 4.5 Air: Enterprise-Ready Innovation

ZAI’s GLM-4.5 (355B total, 32B active) and GLM-4.5 Air (106B total, 12B active), released on July 28, 2025, excel in reasoning and tool-calling (90.6% success rate). Licensed under MIT, they support streaming output and context caching, ideal for enterprise workflows. GLM-4.5’s PowerPoint creation capability highlights versatility, absent in US models’ narrow focus. Their performance, ranking third behind OpenAI’s o3, questions the US’s proprietary superiority.

The US Approach: A Critical Analysis

The US’s reliance on proprietary models, led by OpenAI and Anthropic, prioritizes profit over innovation, stifling progress. Let’s dissect its shortcomings.

OpenAI’s Study Mode: A Superficial Enhancement

OpenAI’s Study Mode offers interactive learning via step-by-step guidance. While useful for education, it’s locked behind Plus, Pro, and Team subscriptions, limiting access. Its reliance on proprietary infrastructure prevents customization, unlike Qwen3’s open-weight flexibility. OpenAI’s delay in releasing an open-source model, possibly to counter Chinese advancements, reflects a defensive posture. This gatekeeping alienates developers, ceding ground to China’s accessible models.

Anthropic’s Claude Max Pricing: Enterprise Exclusion

Anthropic’s Claude Max Pricing targets enterprises with Claude 4 Opus and Sonnet, which excel in reasoning and coding. However, strict rate limits and high costs restrict access, frustrating developers. Claude’s closed ecosystem, while reliable, lacks the customization of Qwen3 or GLM-4.5. This proprietary focus prioritizes revenue over ecosystem growth, contrasting with China’s collaborative approach.

[embed]

Proprietary Lock-In: A Strategic Misstep

The US’s proprietary models depend on cloud-based APIs, locking users into expensive subscriptions. For example, ChatGPT’s $20/month Plus plan and Claude’s tiered pricing dwarf Qwen3’s $1–6/M input costs. This financial barrier excludes startups and researchers, who flock to China’s free, open-weight models. The US’s refusal to embrace open-source risks isolating its developer community, as seen in the rapid adoption of Kimi K2 on platforms like Hugging Face.

Limited Scalability and Context

US models, capped at 128K tokens (e.g., ChatGPT, Claude), lag behind Qwen3-Coder’s 1M-token context window. This restricts applications like large-scale code analysis or long-horizon planning, where Chinese models excel. The US’s reliance on techniques like YaRN to extend context pales against China’s native scalability, highlighting a technical shortfall.

Innovation Stifled by Secrecy

OpenAI and Anthropic’s closed development processes obscure model architectures, hindering academic and community contributions. In contrast, China’s open-weight models invite scrutiny and improvement, accelerating iteration. For instance, Qwen3’s public training data (36T tokens) and MoE details foster collaboration, while OpenAI’s opaque approach breeds mistrust. This secrecy slows US progress, as Chinese labs leverage global feedback to refine models like GLM-4.5.

Technical Comparison: US Falls Short

Architecture and Efficiency

Chinese MoE models, like Qwen3–235B-A22B (22B active) and Kimi K2 (32B active), balance performance and cost, activating only necessary parameters. US models, using dense architectures, demand more resources, increasing latency and expense. Qwen3’s 128 experts, with 8 active per token, achieve efficiency unmatched by Claude’s cloud-heavy design.

Licensing and Accessibility

China’s Apache 2.0 and MIT licenses enable local deployment and commercial use, empowering enterprises to control data and costs. Kimi K2’s modified MIT is less restrictive than OpenAI’s API-only access, which demands constant connectivity. This accessibility gap drives adoption of Chinese models, as developers reject US vendor lock-in.

Benchmark Performance

Qwen3–235B-A22B-Thinking-2507 (92.3 on AIME25) and GLM-4.5 (third overall) rival OpenAI’s o4-mini (92.7) and Claude Opus 4. Qwen3-Coder’s 70% SWE-Bench score matches Claude Sonnet 4, but its open weights allow customization. US models lead in multimodal tasks, but their narrow focus and costs limit versatility compared to China’s broad, accessible offerings.

Ecosystem Impact

China’s models integrate with tools like vLLM, SGLang, and Ollama, supporting diverse hardware. US models, tied to proprietary clouds, exclude users with limited infrastructure. This ecosystem disparity amplifies China’s reach, as developers globally adopt Qwen3 and Kimi K2 for cost-effective innovation.

Consequences of the US Approach

Economic Disadvantage

The US’s high-cost models deter small businesses and startups, who turn to China’s free, open-weight alternatives. For example, Qwen3’s low pricing ($5–60/M output) undercuts Claude, enabling broader adoption. This economic barrier risks shrinking the US’s AI market share, as enterprises prioritize cost over brand loyalty.

Innovation Lag

By withholding open-source models, the US stifles collaborative innovation. China’s transparent approach, sharing weights and training details, accelerates global progress. The US’s secrecy, as seen in OpenAI’s delayed open-source plans, isolates it from community-driven advancements, allowing China to close the performance gap.

Geopolitical Vulnerability

US restrictions on chip exports and business with Chinese firms like ZAI aim to slow China’s AI growth but backfire by spurring domestic innovation. Models like Qwen3 and GLM-4.5, developed despite sanctions, showcase China’s resilience. The US’s proprietary focus leaves it vulnerable to China’s growing influence in global AI adoption.

Developer Alienation

Developers increasingly favor China’s open-source models for their flexibility and cost. Platforms like Hugging Face and LMSYS show surging downloads of Qwen3 and Kimi K2, while US models face criticism for rate limits and costs. This shift erodes the US’s developer base, critical for sustaining AI leadership.

Recommendations for the US

To counter China’s momentum, the US must rethink its approach:

1. Embrace Open-Source: Release permissive-weight models, like a scaled-down GPT-4, to compete with Qwen3 and Kimi K2. This would foster developer trust and ecosystem growth.
2. Reduce Costs: Lower subscription fees or offer freemium tiers to rival China’s pricing, ensuring broader access for startups and researchers.
3. Enhance Scalability: Develop models with larger context windows and hardware-agnostic deployment, matching Qwen3-Coder’s 1M-token capability.
4. Foster Transparency: Share architectural details and training data to encourage academic collaboration, countering China’s open approach.
5. Balance Regulation and Innovation: Ease chip export restrictions strategically to maintain global competitiveness without fueling adversaries.

Conclusion

The US’s proprietary AI approach, exemplified by OpenAI’s Study Mode and Anthropic’s Claude Max Pricing, prioritizes profit over progress, risking its leadership. China’s open-source models — Kimi K2, Qwen3, GLM-4.5, and others — offer performance, accessibility, and cost-effectiveness, exposing US shortcomings. Small choices, like licensing or context length, create significant gaps in this race. To regain momentum, the US must embrace open-source, reduce costs, and foster collaboration.