China’s AI scene is booming with cutting-edge models that challenge the global status quo. Let’s break down the key players, their strengths, and whether they might overtake Western leaders like OpenAI’s ChatGPT or Anthropic’s Claude in the coming years.

DeepSeek: Cost-Effective Brilliance in AI

What it is: DeepSeek is a high-performing large language model (LLM) from China that stunned the world with its low-cost approach to AI reasoning.

Strengths & Influence:

• Performance Parity: Remarkably, DeepSeek’s latest model (V3) “outperforms other open-source models” and matches leading closed-source models. It even beats OpenAI’s GPT-4o on some benchmarks.
• Open-Source & Low-Cost: Unlike ChatGPT, DeepSeek’s models are open source, enabling broad use and customization. They achieve comparable results using less-advanced chips and innovative techniques like the Mixture-of-Experts (MoE), which dramatically reduces training costs. DeepSeek spent just $5.6 million to train V3, versus an estimated $78 million for OpenAI.
• Market Impact: DeepSeek triggered a $1 trillion sell-off in global markets by showing that affordable AI can match big tech’s pricey models. In China, dozens of companies are integrating DeepSeek models, and the DeepSeek chatbot briefly topped China’s App Store, rivaling ChatGPT’s popularity.

Domestic vs Global Focus: DeepSeek has soared within China’s domestic market, gaining government and industry backing. It maintains a low profile internationally, partially due to initial regulatory scrutiny over data privacy.

Innovation & Scalability: DeepSeek exemplifies innovation under constraints. Facing chip bans, it closed the gap with world-class models by smart engineering and a flat, research-centric culture. If it can keep iterating (R2 is already in the pipeline), DeepSeek may continue leveling the playing field globally, spurring rivals to speed up their own efforts.

Prospects: DeepSeek’s trajectory suggests it could become a formidable contender. With proven cost efficiency, strong domestic adoption, and performance approaching OpenAI’s best, it’s a top candidate to challenge ChatGPT in the future.

Alibaba’s Qwen: Open-Source Excellence

What it is: Qwen (short for Tongyi Qianwen, meaning “a thousand questions”) is Alibaba’s ambitious open-source AI model family. Qwen showcases Alibaba’s rising AI prowess, offering models for text, reasoning, and even multimodal tasks.

Strengths & Influence:

• Leading Open Models: Alibaba’s Qwen models power many top open-source LLMs globally. The Qwen team has released advanced models like Qwen2.5 and QwQ-32B, fine-tuned for reasoning with 32 billion parameters. They’ve even rolled out models with 1 million-token context windows – a groundbreaking capacity for long documents.
• Competitive Edge: Qwen-with-Questions (QwQ) was Alibaba’s answer to OpenAI’s reasoning model. It launched matching OpenAI’s early reasoning performance, particularly in math and coding tasks. Recent Qwen models have shown efficiency gains that allow a 32B model to rival much larger models like DeepSeek-R1 (671B). This points to smart scaling and reinforcement learning (RL) optimization.
• Market Impact: By open-sourcing under a permissive Apache 2.0 license, Alibaba enables enterprises worldwide to adapt and commercialize Qwen models. This open approach distinguishes it from proprietary systems like ChatGPT, potentially accelerating global AI development by lowering entry barriers.

Strategic Aim: Alibaba optimizes Qwen for global reach, evidenced by multilingual support, open distribution (e.g., on Hugging Face), and tackling tasks like code generation and long-context reasoning that appeal to international developers.

Innovation & Scalability: With a diverse model lineup (text, vision, audio) and constant updates (Qwen2.5, QwQ improvements), Alibaba’s Qwen demonstrates robust scalability. Their million-token context models hint at future long-form understanding breakthroughs.

Prospects: Qwen’s open-source strategy combined with Alibaba’s resources make it a strong long-term player. It directly competes with GPT-style models and could potentially surpass closed models in certain domains by harnessing the global developer community’s innovations.

Tencent’s Yuanbao: AI in Your Pocket

What it is: Yuanbao is Tencent’s conversational AI assistant, integrated into WeChat (China’s ubiquitous super-app). It’s backed by Tencent’s Hunyuan model and DeepSeek’s reasoning, blending in-house AI with startup innovation.

Strengths & Influence:

• Massive User Base: By launching inside WeChat (with over a billion users), Yuanbao achieved instant market penetration. Tencent even added a one-click download in WeChat, making Yuanbao extremely accessible. This user integration catapulted Yuanbao to #1 on China’s iOS App Store, overtaking DeepSeek’s app.
• Tech Blend: Yuanbao leverages Tencent’s own Hunyuan LLM (a competitor to models like GPT-4) and augments it with DeepSeek’s R1 reasoning. This hybrid approach ensures strong conversational abilities with fast response times (Tencent also rolled out a “Turbo S” model for speed).
• Competitive Momentum: Tencent has been catching up after a late start. Yuanbao’s surge in popularity shows Tencent’s strategy—deep platform integration—is working. Tencent is also reportedly investing heavily in AI hardware (Nvidia H20 chips) to boost Yuanbao’s capabilities further.

Domestic vs Global: So far, Yuanbao is primarily domestic, thriving within China’s WeChat ecosystem. While WeChat has overseas users, Yuanbao’s content likely aligns with Chinese regulations (e.g., avoiding sensitive topics), which might limit global adoption unless tailored for other markets.

Innovation & Scalability: Tencent’s focus is on user experience and ecosystem synergy. By embedding AI in everyday apps, it gains vast data for improvement. Yuanbao’s model can scale by continuous integration with Tencent’s services (e.g., QQ, cloud gaming), though the innovation seems application-driven rather than foundational research.

Prospects: Yuanbao’s path to challenge ChatGPT or Claude globally would require Tencent to push it beyond WeChat – possibly via an English version or integration in popular games. Given Tencent’s clout, Yuanbao could loom large in consumer AI, especially in Asia. For now, it’s a domestic heavyweight with potential for global play if Tencent chooses.

Baidu’s ERNIE 4.5 & X1: The Accessible AI Duo

What they are: Baidu’s ERNIE (Enhanced Representation through kNowledge IntEgration) is a family of LLMs. ERNIE 4.0 was Baidu’s answer to GPT-4, and ERNIE 4.5 likely represents an iterative upgrade. X1, possibly referring to ERNIE Bot X, is geared toward cost efficiency and accessibility, perhaps as a lighter model for enterprises or public use.

Strengths & Influence:

• Multimodal & Multilingual: ERNIE 4.0 boasted abilities beyond text – such as images and videos – in its responses, aligning with GPT-4’s multimodal vision. It excels in understanding Chinese context (no surprise, given Baidu’s search engine roots).
• Claimed Parity with GPT-4: Baidu claimed ERNIE 4.0 “rivals models such as GPT-4”. While independent benchmarks have been limited, Baidu has showcased ERNIE answering real-time and culturally specific queries better than GPT-4 in some cases.
• X1 for Accessibility: Baidu has emphasized making AI widely accessible. If “X1” refers to a variant aimed at broader availability (e.g., cloud API or low-cost version), it aligns with Baidu’s strategy to embed ERNIE across its products (search, cloud services, smart devices).

Domestic vs Global: Baidu’s AI strategy is somewhat China-focused. ERNIE powers Baidu’s own ecosystem (Baidu Search, Xiaodu smart speakers), and content moderation is tuned to Chinese regulations. Globally, Baidu might leverage ERNIE via open APIs or partnerships, but it hasn’t seen the same international uptake as ChatGPT.

Innovation & Scalability: Baidu’s strength is its massive data from search and services, which can feed ERNIE’s training (especially in Chinese language). Cost efficiency is highlighted by mentions of ERNIE being optimized for cheaper, faster deployment. Baidu also introduced ERNIE 4.0 in search, blending LLM with search results for an enhanced Q&A experience. Scalability is proven by Baidu’s rapid upgrades (if ERNIE 4.5 is out shortly after 4.0, it indicates fast iteration).

Prospects: Baidu has the AI chops and data to keep ERNIE competitive. However, to surpass OpenAI/Anthropic, Baidu must engage the global community or deliver breakthrough features. If ERNIE’s English and coding skills catch up, and if Baidu offers it as an open platform, it could become a top contender, especially in Asia-Pacific markets.

Moonshot AI: Ambitious General AI Startup

What it is: Moonshot AI is a startup gunning for general AI. Founded in 2023 (amid the ChatGPT wave), it has quickly built a high profile and impressive tech. Notably, Moonshot developed Kimi Chat, known for an extremely long context window.

Strengths & Influence:

• Long-Context Pioneer: Moonshot’s Kimi Chat could process 200,000 Chinese characters in one input – likely the longest context capacity among commercial models at the time​. This is a huge leap (for reference, GPT-4’s context is 32K tokens, roughly 24K English words). This ability lets Kimi handle lengthy documents or multi-turn dialogues without losing track.
• Rapid Traction: By early 2024, Kimi Chat soared in popularity, trending on social media and ranking 3rd among China’s AI products by user traffic. Moonshot reached a $3 billion valuation in just a year, reflecting investor confidence.
• Technical Roots: Founder Yang Zhilin is a Tsinghua and Carnegie Mellon alum, with deep AI research experience​. Moonshot emphasizes research breakthroughs (like scaling laws, per its founder’s talks), hinting it’s chasing fundamental AI advances, not just applications.

Domestic vs Global: Moonshot is China-based but “general AI” ambition implies global vision. Yet, like many Chinese firms, it may face export controls or scrutiny if expanding abroad. Domestically, it competes with the “AI Tigers” (top startups), but globally it’s still under the radar compared to OpenAI.

Innovation & Scalability: Moonshot’s focus on long context and presumably efficient scaling tackles key AI challenges (memory and consistency over long inputs). If they keep innovating on model architectures, they could leapfrog existing models on tasks requiring reading books or lengthy reports. Scalability will depend on computing power and data; the legal hiccups with investors may be a distraction, but also show the intensity around this startup.

Prospects: While Moonshot is in early days, its bold technical bets (like long context) could pay off. If they crack more general AI problems and avoid regulatory pitfalls, Moonshot AI might produce the next breakthrough model – perhaps not immediately surpassing ChatGPT, but carving out a leading niche (like best long-document AI). Keep an eye on it for surprises.

MiniMax: Multimodal “AI Tiger” on the Rise

What it is: MiniMax is a well-funded Shanghai-based AI startup (valued ~$2.5B) specializing in multimodal AI – models that handle text, vision, and audio together.

Strengths & Influence:

• Diverse Models: In early 2025, MiniMax unveiled three models: MiniMax-Text-01 (text LLM), MiniMax-VL-01 (vision+text multimodal), and T2A-01-HD (text-to-audio, generating speech). This trifecta covers a broad spectrum of AI tasks, comparable to combining ChatGPT (text), GPT-4V (vision), and a speech generator.
• Competitive Performance: MiniMax claims its 456B-parameter Text-01 model outperforms Google’s Gemini 2.0 Flash on certain benchmarks (MMLU for math, etc.). It also states the VL-01 model rivals Anthropic’s Claude 3.5 on multimodal tasks – though still behind GPT-4 on some tests. The willingness to compare directly with top Western models shows MiniMax’s confidence.
• Massive Context Window: MiniMax-Text-01 has an extremely large context window of 4 million tokens​ (yes, million!). That’s orders of magnitude above current Western models, enabling analysis of around 3 million words in one go​. This could be a game-changer for tasks like whole-dataset analysis or lengthy conversation continuity.

Strategic Position: Backed by Alibaba and Tencent, MiniMax has strong support. It’s building open-source models, signaling a strategy to influence the developer community and reduce reliance on foreign tech. Its multimodal expertise fits into trends where AI systems need to see, talk, and understand context all at once – ideal for smart assistants, content creation, etc.

Innovation & Scalability: MiniMax is clearly innovating on scale (456B parameters, huge context) and breadth (multimodality). Scaling such large models is costly, but its big funding helps. If the models are open-source as hinted, they could gain widespread adoption and community-driven improvements, aiding scalability through collective efforts.

Prospects: By targeting the high end of tech (rivaling GPT-4/Claude) and keeping models open, MiniMax could become a major AI platform. Surpassing OpenAI or Anthropic won’t be easy, but if its models consistently prove better on key tasks (and are freely available), it might tip the scales. Watch its multimodal model – as AI moves beyond text, MiniMax wants to lead that pack.

Huawei’s Pangu Model: Industry Specialist

What it is: Huawei’s Pangu is a suite of AI models, each tailored for specific industries – finance, meteorology, mining, etc. – rather than a single general chatbot. It’s China’s first series of commercial foundation models built for enterprise solutions.

Strengths & Influence:

• Industry-Leading Applications: Pangu-Weather is famed for being the first AI to outperform traditional numerical weather prediction, enabling 10,000× faster forecasts (a Nature-published feat). Pangu-Mining aims to streamline coal mining operations through AI. There’s also Pangu-Finance, Pangu-Government, etc., each trained on domain-specific data. This specialization yields expert-level results in each field, which general models like ChatGPT can’t match in depth.
• Hierarchical & Scalable: Pangu has a three-layer architecture: core LLMs (L0), industry models (L1), and scenario-specific models (L2). This modular design means businesses can take base models and quickly train them for niche tasks. Huawei also offers Pangu in different sizes (1B, 10B, 100B+ parameters) to balance cost and power. This flexibility is a big plus for real-world deployment.
• Global Footprint via Partnerships: Despite U.S. sanctions, Huawei found ways to push Pangu globally: e.g., partnering with Europe’s ECMWF for weather forecasts, releasing Pangu-Weather for free public use, and showcasing Fintech solutions on global stages. This hints Huawei can sidestep some restrictions by focusing on beneficial, non-controversial uses (like weather and healthcare).

Domestic vs Global: Huawei, being on export control lists, has hurdles globally. But Pangu’s industry focus may help – companies worldwide might adopt a top-notch weather model or finance model if available. Domestically, Pangu is heavily promoted to modernize industries (mining, etc.), aligning with China’s national tech self-reliance goals.

Innovation & Scalability: Huawei’s advantage is its hardware-software synergy (it makes AI chips like Ascend) and R&D might. Pangu’s continuous updates (now at version 5.0) show Huawei’s commitment. They innovate by combining AI with domain knowledge (e.g., physics for weather). Scalability is achieved by that layered approach – they don’t need one model to do everything, they need many models doing specific things very well.

Prospects: Pangu might not directly compete with ChatGPT in casual conversation, but in its domains, it could outshine general models. In a way, Huawei is creating a parallel AI ecosystem: specialized models that collectively could match the versatility of one giant model. If enterprise AI adoption grows, Pangu could be as valuable as a GPT – just in a different, more practical way. Surpassing OpenAI/Anthropic overall is unlikely unless general-purpose Pangu variants emerge, but in enterprise AI leadership, Pangu is a strong contender.

SenseTime: From Vision to Generative AI

What it is: SenseTime is a long-time Chinese AI leader known for computer vision (e.g., facial recognition). Recently, it shifted towards generative AI, launching the SenseNova LLM series (and SenseChat chatbot).

Strengths & Influence:

• Deep Expertise in AI: SenseTime has a decade of AI research behind it (though mostly vision). It leveraged this to claim SenseNova has capabilities “comparable to OpenAI’s latest GPT”. Its SenseChat V5 reportedly outperforms GPT-4 Turbo on benchmarks like MMLU, HumanEval – bold claims that indicate serious optimization, at least for Chinese context and coding.
• Multimodal Prowess: SenseTime’s background means it can integrate vision into chat. They showcased a real-time multimodal model in SenseNova 5.0, aiming for on-par performance with GPT-4’s image understanding. Few companies have both top-tier vision and language tech; SenseTime is one of them.
• Applications & Adoption: SenseTime’s AI has wide applications in China (smart cities, retail, healthcare). As it pivots to generative AI, it can embed SenseNova into these verticals – e.g., smarter surveillance analytics, AI assistants in healthcare diagnostics, etc. Also, its public profile (Hong Kong-listed) means it’s somewhat transparent and used to global collaboration, which might help in international markets.

Strategic Focus: Initially domestic, SenseTime has faced U.S. sanctions (for alleged misuse of its tech in surveillance). This forced it to double down on China and friendly markets. It’s reorganizing to chase generative AI growth and profitability, essentially trying to reinvent itself as China’s OpenAI.

Innovation & Scalability: The challenge for SenseTime is scaling LLMs as effectively as it did vision. They’re investing heavily (despite being loss-making) to catch up. SenseNova’s iterations (5.0 to 5.5) in a short time show fast scaling. Its WYSIWYG model mentioned in analyses suggests innovative UIs or tools for content generation.

Prospects: SenseTime has the talent and government backing to be a contender. To surpass ChatGPT, it needs to prove its models not only match but outperform consistently, in multiple languages, and innovate (maybe via unique vision-language features). If it can shake off the sanction shackles and partner globally, SenseTime could become a strong third pillar (with OpenAI and Google). In the near term, it’s a rejuvenated player to watch, especially in Asia’s AI markets.

Zhipu AI: The Academic Trailblazer

What it is: Zhipu AI, born out of Tsinghua University, is one of China’s earliest and most research-driven AI startups (founded in 2019). It’s known for the ChatGLM series, bilingual chat models open-sourced to the community.

Strengths & Influence:

• Bilingual Foundation Models: Zhipu co-developed GLM-130B, one of the first open bilingual (Chinese-English) LLMs, and ChatGLM-6B/ChatGLM2 as ChatGPT alternatives. These open models gave Chinese developers a homegrown chatbot to experiment with, boosting local innovation.
• Latest Tech – GLM-4-Plus: In Aug 2024, Zhipu introduced GLM-4-Plus, claiming it “performs on par with OpenAI’s GPT-4o”. That’s an impressive stride for a non-tech-giant. They also released GLM-4-Voice, enabling real-time voice conversations in Chinese/English with human-like intonation. This shows Zhipu’s focus on a well-rounded AI (text+voice).
• Academic and Industry Synergy: Being Tsinghua-affiliated means Zhipu has strong research ties. It likely has access to top talent (students, professors) and can test ideas quickly. It’s also attracted investments from Alibaba, Tencent, and state entities, bridging academia and industry in a unique way.

Domestic vs Global: Zhipu is oriented towards open platforms (its BigModel site is a top repository in China). This openness could pave the way for global collaboration, as code and models are accessible. However, recent U.S. trade restrictions (added to a restricted list in 2024) could hinder direct expansion or compute resources.

Innovation & Scalability: Zhipu’s innovation lies in efficiency and openness. GLM models often use clever training tricks (e.g., INT4 quantization in GLM-130B) to run large models with limited hardware. Scalability might be challenging due to fewer resources than tech giants, but partnerships and its billion-yuan funding round will help.

Prospects: If any Chinese model is to surpass ChatGPT on quality in an open-source manner, Zhipu is a top candidate. It mixes academic rigor with industry pragmatism. The key will be maintaining a rapid pace of improvement (to truly match GPT-4’s ever-evolving versions) and navigating geopolitical headwinds. If successful, Zhipu’s open models might become the go-to ChatGPT alternative worldwide, not just in China.

Baichuan Intelligence: The New Wave in LLMs

What it is: Baichuan Intelligence is a startup led by Wang Xiaochuan (ex-CEO of Sogou, a major search engine). It’s a rising star known for open-source LLMs like Baichuan-13B, aiming to rival OpenAI’s offerings​.

Strengths & Influence:

• Stellar Leadership: Wang Xiaochuan’s background as a search pioneer and Tsinghua prodigy brings credibility​. His vision: “China needs its own OpenAI.” Baichuan quickly got $50M funding and a team of 50 within months​, showcasing momentum.
• Open-Source Large Models: Baichuan-13B is open, commercial-friendly, and trained on 1.4 trillion tokens​ (40% more data than Meta’s LLaMa-13B​). This heavy training paid off with Baichuan-13B reportedly outperforming LLaMa-13B by a large margin​ and being optimized for practical use. Baichuan quickly followed with a 7B model and even a larger 53B model​, showing an aggressive release cycle.
• Focus Areas: Baichuan is also looking at math and healthcare AI (perhaps applying LLMs in those fields)​, diversifying beyond generic chat. They even launched a financial-specialized model (Baichuan4-Finance)​, indicating an industry-tailored approach similar to Huawei’s Pangu but for software.

Domestic vs Global: Given its open-source nature and English-Chinese training, Baichuan is positioned to be globally relevant. It’s still new (founded 2023), so it hasn’t made a global splash yet beyond tech circles. Domestically, it’s one of the celebrated “AI Tigers” startups, often mentioned alongside DeepSeek, etc.

Innovation & Scalability: Baichuan’s rapid development from 7B to 13B to 53B models in about a year​ is impressive. They emphasize quality data and balanced training (bilingual). Scalability might be a concern without a giant parent company, but Wang’s reputation likely helps in attracting talent and capital. Their pivot to some closed-source for the largest model​ shows they balance open community benefits with proprietary advantages for revenue.

Prospects: Baichuan Intelligence could be the dark horse in this race. If their models continue to improve so quickly, and remain open for broad use, they could accumulate a strong user base. Surpassing OpenAI is a tall order, but Baichuan could at least become China’s local OpenAI equivalent, and a key contributor to the global open model ecosystem. If one day Baichuan-53B or future variants beat GPT-4 in benchmarks and are freely available, that would indeed put OpenAI on notice.

Regulatory & Ethical Considerations in Chinese AI

A critical factor in global competition is how government policies and ethics shape AI development:

• Government Support & Oversight: The Chinese government actively supports AI breakthroughs (financially and via initiatives) but also imposes strict content regulations. Models like DeepSeek and Yuanbao refuse sensitive prompts (e.g., on politics), aligning with censorship rules. This could hinder global appeal where users expect free-form Q&A, but it’s necessary for any model operating in China.
  
• Export Controls: U.S. sanctions on advanced chips forced Chinese AI labs to innovate with what they have. While this led to creativity (DeepSeek’s methods), it also means scaling up is harder. Companies like SenseTime and Zhipu being put on trade blacklists shows geopolitics can slow their global reach and collaboration.
  
• Ethical AI & Bias: Chinese models may carry different biases, having been trained on Chinese web data and following local norms. Ethically, they avoid certain topics, which could be seen as either a drawback (less open discourse) or a safeguard (less risk of offensive or disallowed content). For global competitiveness, they’ll need to prove they can handle ethics and bias as well as Western models, but in a culturally adaptable way – perhaps tuning models per region.
  
• Data Privacy: China’s data laws (like the Personal Information Protection Law) are strict in their own way. Startups like DeepSeek had privacy practice questions (government told it to keep a low profile due to data concerns). To go global, Chinese AI firms must assure users and regulators worldwide that data is handled responsibly and not overly influenced by state interests.
  

Net Effect: Chinese AI models have strong state backing and fewer market-driven constraints (they can focus on research over immediate profit, e.g., DeepSeek’s lab-like culture). But they also operate within regulatory guardrails that might slow certain features or global adoption. However, with alignment to ethical standards and transparency, these models can still compete internationally – especially if they excel technically.

Future Outlook: Will Chinese Models Surpass ChatGPT or Claude?

Synthesizing all insights:

• Current Standing: Chinese models have rapidly caught up. DeepSeek V3 challenging GPT-4o, MiniMax and Zhipu making GPT-4 comparisons, and Baichuan open-sourcing competitive models, all indicate China is nearly on par in AI quality. In China’s domestic market, ChatGPT is banned, so these models already “surpassed” it in usage. Globally, none have dethroned ChatGPT yet, but the gap is closing.
  
• Strengths to Leverage: Cost-efficiency (DeepSeek), open-source community (Qwen, Baichuan, Zhipu, MiniMax), massive user integration (Yuanbao), and domain specialization (Pangu) are edges Chinese models hold. OpenAI and Anthropic should watch these as they could erode the Western lead on multiple fronts: price, openness, user base, and vertical performance.
  
• Challenges: To surpass OpenAI or Anthropic, Chinese models must excel not just at one thing, but across the board: reasoning, creativity, coding, multilingual fluency, safety, etc. They also need to gain trust and adoption outside China. That might involve localizing models for different cultures and proving they can handle content freely (where allowed).
  
• Global AI Competition: It’s not zero-sum. We might see a world where OpenAI’s models and top Chinese models compete and coexist. Users might pick based on needs: e.g., a developer might fine-tune Alibaba’s Qwen for their app (because it’s open) but use ChatGPT for brainstorming. However, if one of these Chinese models makes a breakthrough – say, achieving AGI-like reasoning or far superior efficiency – it could leapfrog the rest.
  

Bottom Line: In the near future (1-2 years), Chinese AI models will significantly narrow the gap, and on some specific benchmarks or applications, they may outperform ChatGPT or Claude. Surpassing them in general capability might take longer, but it’s increasingly plausible given the momentum. With heavy investments, smart talent, and a supportive (if controlled) environment, China’s AI models have the potential to become the next major contenders on the world stage. OpenAI and Anthropic no longer compete in a vacuum – the dragon has awakened in the AI race.