AI models

Are the Latest Foreign “Self-Developed” AI Models Just Rebranded Chinese Ones?

Honghao Wang

— 4 min read
Are the Latest Foreign “Self-Developed” AI Models Just Rebranded Chinese Ones?

Foreign Developers: Should We Start Learning Chinese?

It’s a curious sight — the latest large model from a U.S. tech company intermittently “thinks aloud” in Chinese during its reasoning process.

image

---

Cursor 2.0: New Model and Multi-Agent Collaboration

This week, popular AI coding tool Cursor rolled out its 2.0 update, featuring:

  • Composer — Cursor’s first in-house code model.
  • A new interface enabling parallel collaboration among multiple AI agents.

Key Takeaways

  • Composer is a Mixture-of-Experts (MoE) model trained with reinforcement learning.
  • Built for real-world coding tasks with exceptional speed.
  • Achieves 4× faster output than comparable models.
image
image

Origins and Suspicion

Despite the excitement, developers noticed Composer frequently outputs intermediate reasoning in Chinese — sparking speculation it may be based on Qwen Code.

Cursor’s blog revealed that Composer evolved from a prototype agent named Cheetah, designed to study ultra-fast agents. Composer is essentially Cheetah’s smarter, faster successor.

image

---

Cognition SWE-1.5: Another Surprise

AI startup Cognition also introduced SWE-1.5, a high-speed coding agent model boasting:

  • Tens of billions of parameters
  • Up to 6× faster than Haiku 4.5
  • 13× faster than Sonnet 4.5
image

Released on their IDE platform Windsurf:

image

Community Investigation

  • Developer “jailbreak” testing revealed SWE-1.5 may be derived from GLM developed by Chinese AI company Zhipu AI.
  • Zhipu’s official X account even retweeted congratulations.
image
image

---

Industry Commentary: Evidence of Chinese Base Models

Several AI experts believe both Composer and SWE-1.5 are fine-tuned versions of Chinese open-source models:

  • @deedydas suggested SWE-1.5 is a customized GLM-4.6 running on Cerebras hardware.
  • Composer’s output style matches the distinct “Chinese-style” reasoning traces.
image

Cerebras later announced they will launch zai-glm-4.6 coding model — effectively confirming the connection.

Expert Analysis

Daniel Jeffries argues:

  • Both Windsurf and Cursor likely didn’t train these models from scratch.
  • Fine-tuning + RL training is far cheaper and uses readily available coding datasets.
  • Building a base model independently requires massive funding, data, and infrastructure — beyond the reach of many startups.
image

---

Open-Source + Fine-Tuning: The Emerging Development Playbook

This approach — leveraging strong open-source base models then enhancing them with domain-specific fine-tuning and reinforcement learning — is becoming standard.

Benefits

  • Lowers entry barriers for small teams
  • Rapid route to near-SOTA performance
  • Aligns with global open-source innovation trends

Example:

AiToEarn官网 — an open-source global AI content monetization platform that enables creators to:

  • Generate AI-powered content
  • Publish simultaneously to platforms (Douyin, Kwai, WeChat, Bilibili, Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X)
  • Analyze and monetize their work

Its AI模型排名 tool (link) helps identify the best model for specific needs — echoing the fine-tuning philosophy driving products like Cursor and Windsurf.

---

Funding and Open-Source Advocacy

Jeffries believes:

  • Cursor and Windsurf teams lack the infrastructure to train truly foundational models from scratch.
  • Many large labs have already achieved massive scale, making it nearly impossible for smaller firms to compete without leveraging open source.
  • Some stakeholders opposing open source hinder modern software innovation.
image
image

His post, reshared by Yann LeCun, has fueled broad discussion on the role of Chinese open-source models in powering Western products.

image

Humor from the community: “Is it time to start learning Chinese?”

image

---

Chinese Open-Source Models Are Going Global

On October 29, during NVIDIA’s GTC conference in Washington, CEO Jensen Huang emphasized:

  • Open-source models have become critical drivers of AI application speed.
  • The global AI community — from researchers to enterprises — needs open source.

Market Leadership

  • Since 2025, Alibaba’s Qwen dominates the open-source model market share.
  • Qwen leads in derivative model count worldwide.
image

Open-source models now excel in:

  • Reasoning ability
  • Multimodal capabilities
  • Specialized domain expertise

For startups like Cursor and Cognition, this may be the foundation of their success.

---

Benchmark Rankings Show Chinese Dominance

On HuggingFace trending lists, most top models come from Chinese companies:

  • MiniMax
  • DeepSeek
  • Kimi
  • Baidu HunYuan
  • Qwen
  • Meituan LongCat
image

Similarly, ArtificialAnalysis ranks many Chinese-developed models at the top, considering:

  • Performance
  • Speed
  • Context window size
  • Parameter count
  • License type
image

---

The Competitive Shift

The technical sophistication, developer adoption, and capabilities of Chinese open-source large models are reshaping the global AI competitive landscape — with leadership roles gradually changing.

---

AiToEarn: Empowering Creators in the New AI Era

Platforms like AiToEarn官网 align perfectly with this shift, giving AI creators:

  • AI-driven content generation tools
  • Cross-platform publishing
  • Analytics and model rankings (link)
  • Sustainable monetization options

This mirrors the growing reliance on fine-tuned open-source models, and suggests that for tech teams and creators alike, learning Chinese might soon be more than just a fun idea — it could be a competitive advantage.

Read more