The best Chinese neural networks: who surpassed ChatGPT and Claude

This review is current as of July 2026. The industry is evolving rapidly—new model versions are released almost monthly, so specific figures and ranking positions should be verified against official developer documentation.

DeepSeek — Flagship for Text and Reasoning

If 2023 was defined by ChatGPT, then 2025–2026 are defined by DeepSeek. The company demonstrated that it is possible to create a model on par with GPT-4o while spending tens of times less on training. Current models in the lineup include DeepSeek-V3 and R1, built on a Mixture of Experts (MoE) architecture, which means only a small portion of the network is activated per request. This significantly reduces computational costs without compromising quality.

Strengths: logical reasoning, structured responses, document processing, and basic coding. The model confidently understands Russian.

Qwen/Alibaba Cloud — The Broadest Ecosystem

The Qwen ecosystem is one of the broadest on the market, with separate lines for code, vision, and omnimodality.

The latest proprietary flagship is Qwen3.7-Max, an “agent-era model” optimized for autonomous task execution involving hundreds or thousands of steps. In a demonstration, it autonomously optimized code for 35 hours.

Open models are also available: the flagship Qwen3.6-Plus was released on April 2, 2026, and the open-weight version Qwen3.6-35B-A3B on April 16. This is an MoE model with 35 billion total parameters and only 3 billion active per request, allowing it to run quickly even on relatively modest hardware and locally via Ollama or LM Studio.

GLM/Zhipu AI

Zhipu AI is a spin-off of Tsinghua University and the first Chinese company developing large language models to be placed under U.S. sanctions in January 2025. In January 2026, the company conducted an IPO on the Hong Kong Stock Exchange.

The current flagship is GLM-5.2: an open model under the MIT license with 744 billion parameters, of which 40 billion are active, and a context window of one million tokens.

Notably, its predecessor, GLM-5, was trained on 28.5 trillion tokens entirely on Chinese Huawei Ascend accelerators without using NVIDIA hardware.

The company positions the model as a Claude-level solution for programming at a significantly lower API cost. A separate model for autonomous agents is GLM-5.1, which can work on a single task without human intervention for up to eight hours straight: writing code, running tests, compiling projects, and refining architecture.

Kimi/Moonshot AI — Neural Network with a Large Context Window

Kimi’s main feature is its ability to handle very long texts. The neural network can analyze an entire book, a large report, or extensive source code without needing to split them into parts. Because of this, it is well-suited for document analysis, summarizing key points, searching for specific information, and assisting with programming. In 2026, version K2.6 was also released, featuring improvements for writing and analyzing code.

ERNIE/Baidu — Ommimodal Contender for the Global Top

ERNIE is Baidu’s flagship neural network, capable of processing not only text but also images, audio, and video. It is suitable for writing texts, analyzing information, generating content, and solving everyday work tasks.

In January 2026, the model entered the global top 10 of the LMArena ranking, becoming at that time the only Chinese model in the top ten.

A more recent version, ERNIE 5.1, ranks fourth globally according to LMArena and remains the best result among Chinese models.

Specialized Models for Code and Agents

In April 2026, KAT-Coder-Pro V2 ranked sixth in the global AA Coding Index benchmark and became one of the best specialized models for programmers.

Another notable player in the autonomous agent niche is Step-3.5-Flash. This is an MoE model with 196 billion total parameters, 11 billion active per request, a context window of 256 thousand tokens, and a generation speed of up to 350 tokens per second. It was specifically trained for planning multi-step tasks and working with external tools.

Image and Video Generation

Several Chinese products stand out in the visual segment:

Kling AI — one of the most well-known video generators. It creates realistic videos from text descriptions or images and is used for advertising, social media content, and short films.

Hailuo AI — focuses on natural movements and realistic animation. It is particularly well-suited for bringing photos to life and creating cinematic videos.

PixVerse — allows for the rapid creation of short videos from text or photos. It is often used for content on TikTok, Instagram, and other social networks.

Wan and Qwen-Image — Alibaba models for generating images and videos, integrated into the Qwen ecosystem. They are suitable for creating illustrations, banners, concept art, and other visual content.