Meituan Just Banned ByteDance's AI Model. Here's What It Reveals About China's Internal AI Power Struggle

On July 2, 2026, Meituan — China's $100-billion food delivery and local services giant — sent an internal notice to every business unit in the company. The message was blunt: stop using Doubao (豆包), ByteDance's flagship large language model.

All teams were ordered to audit their existing Doubao integrations and begin migrating to either LongCat (Meituan's own in-house model) or DeepSeek (the open-source model from Hangzhou-based DeepSeek AI). Teams that cannot migrate must submit a detailed justification and obtain executive-level approval.

This is not the first time Meituan has pulled this move. In April 2026, the company stopped recommending Alibaba's Qwen model for internal use. Any team still wanting to use Qwen must now obtain approval from X3-level executives — essentially, C-suite bosses.

Within three months, Meituan has effectively locked out two of China's three major LLM providers — Alibaba and ByteDance — in favor of its own model and DeepSeek.

For foreign observers, this might look like a routine corporate IT decision. It isn't. It is a window into how China's tech giants are quietly reshaping the country's AI supply chain, and what that means for the global AI race.

What Meituan Is Actually Doing

To understand this decision, you need to understand Meituan's position in China's AI landscape.

Meituan is not just a food delivery company. It operates across food delivery, flash shopping, hotel bookings, local services, and bike-sharing — processing tens of millions of orders daily across 3,000+ Chinese cities. Its AI needs are enormous: route optimization, demand forecasting, merchant analytics, customer service, and now an AI assistant called "Xiao Tuan" (小团) embedded directly in the Meituan app.

Since 2023, Meituan has been developing its own LLM, LongCat. The model has gone through 12 iterations covering text, image, voice, and video modalities. In June 2026, Meituan released LongCat-2.0 — a 1.6-trillion parameter Mixture-of-Experts (MoE) model trained entirely on 50,000 domestic Chinese AI accelerators, a first in the industry.

LongCat-2.0's key specs are remarkable:

• 1.6 trillion total parameters, with average activation of 48 billion per inference (dynamic range 33B–56B)
• 30+ trillion tokens of pre-training data (Chinese, English, multilingual, code)
• 1 million token context window
• SWE-Bench Pro score: 59.5% — surpassing GPT-5.5 (58.6%) and Claude Opus 4.6 (57.3%)
• Training cost 31% lower than comparable overseas clusters
• Fully open-sourced under MIT license

This is the model Meituan wants its teams to use. And the alternative it is willing to accept? DeepSeek.

Not ByteDance's Doubao. Not Alibaba's Qwen. DeepSeek.

Why DeepSeek? The Unspoken Logic

Meituan did not publicly explain why DeepSeek made the cut while Doubao and Qwen did not. But the industry logic is not hard to read.

1. DeepSeek Is the "Neutral" Option

Unlike Doubao (ByteDance) or Qwen (Alibaba), DeepSeek is not primarily a platform company competing with Meituan for users, merchants, or advertising dollars. ByteDance owns Douyin (TikTok's Chinese version), a direct competitor to Meituan's local life services. Alibaba operates Ele.me and Flash Shopping, both competing directly with Meituan's core business. Using a competitor's AI model creates strategic dependency risks that go beyond technology.

DeepSeek, by contrast, is a pure-play AI research company. It does not compete with Meituan in any consumer or merchant market. Adopting DeepSeek is a technology decision, not a geopolitical one.

2. DeepSeek's Open-Source Model Is Genuinely Excellent

DeepSeek's models consistently rank among the top performers on coding, reasoning, and safety benchmarks globally. Its open-source availability means Meituan can fine-tune, deploy, and integrate without vendor lock-in. The model is the safe, high-quality fallback when you don't want to bet everything on your own model.

3. The Broader Trend: Chinese Companies Are Choosing DeepSeek Over Domestic Rivals

This pattern is not unique to Meituan. Coinbase made DeepSeek and Zhipu GLM its default engineering models. Airbnb migrated to Alibaba's Qwen for specific workloads. American companies are choosing Chinese models. And now, Chinese companies are choosing DeepSeek over other Chinese models.

China's Tech Giants Are Building Walled Gardens

Meituan's model strategy follows a clear trajectory:

• Self-developed first. LongCat is the default. It powers "Xiao Tuan," merchant assistants, route optimization, and the company's AI IDE tool CatPaw (which has 95%+ internal adoption and generates over 50% of new code).
• DeepSeek as the approved external alternative. For use cases where LongCat is not sufficient or not the right fit.
• Everything else requires executive approval. ByteDance's Doubao and Alibaba's Qwen are effectively blocked.

This is not just an IT consolidation exercise. It is a supply chain sovereignty strategy — the same logic that drives nations to diversify their semiconductor suppliers or energy sources. Meituan is ensuring that its AI infrastructure is not dependent on companies that compete with its core business.

The fact that Meituan trained LongCat-2.0 entirely on domestic AI accelerators — not Nvidia chips — adds another layer. This is a company building end-to-end AI sovereignty: domestic chips, domestic models, domestic training infrastructure.

What This Means for ByteDance

For ByteDance, Meituan's decision is a notable blow.

ByteDance's Doubao is technically impressive. At the recent FORCE conference, ByteDance announced that Doubao's daily token consumption has surpassed 180 trillion — a 10x increase in one year. ByteDance claims a 49.5% market share in China's Model-as-a-Service (MaaS) market, meaning roughly one in every two tokens consumed by Chinese enterprises is processed by ByteDance's infrastructure.

But impressive usage numbers do not guarantee enterprise trust. ByteDance's core business — short video, social media, e-commerce — puts it in direct competition with almost every other major Chinese tech company. Meituan (local services), Alibaba (e-commerce), Tencent (social), and Baidu (search) all have reasons to avoid depending on ByteDance's AI infrastructure.

This is ByteDance's fundamental strategic challenge in the AI model market: the company that owns TikTok's parent cannot easily be a neutral AI infrastructure provider to its competitors.

Meituan's decision signals that even when a model is technically competitive, platform rivalry can override technical merit in enterprise AI procurement.

The Broader Signal: China's AI Industry Is Fragmenting Into Camps

Meituan's three-phase model restriction — first Qwen, now Doubao, keeping only LongCat and DeepSeek — reflects a broader pattern in China's AI industry:

The competition between these camps is not just about model performance. It is about ecosystem control, data sovereignty, and strategic independence — the same forces that shape geopolitical alliances, now playing out in the AI supply chain of a single country.

What Foreign Readers Should Take Away

If you are reading this from outside China, here are the key points:

Conclusion: The AI Cold War Is Internal, Too

When we talk about the "AI race," the narrative usually focuses on US vs. China. But inside China, there is another race happening — between companies that want to own the entire AI stack and those willing to rely on specialized, independent providers.

Meituan's decision to ban ByteDance's Doubao while embracing DeepSeek and its own LongCat is a small data point with a large implication: in the age of AI, supply chain trust matters as much as model performance. And the companies that can position themselves as neutral, high-quality infrastructure providers — like DeepSeek — may end up winning without ever building a consumer product.

The sweeping monk is still sweeping. But now we know who is watching.

Sources: IT之家, 凤凰网科技, Tech星球, 中国经济网