Os 5 melhores modelos LLM chineses em 2026: qual é o melhor para programação?

The Best Chinese LLM Models at a Glance

The independent scores come from the current Artificial Analysis Intelligence Index, which combines nine evaluations covering coding, terminal work, knowledge, mathematics and reasoning. It is broader than a coding-only leaderboard, so it should be treated as one signal rather than a final verdict. See the Artificial Analysis methodology and current model results.

Prices are GPT Proto rates checked on July 20, 2026. They may change.

How We Ranked the Chinese LLM Models

A “best Chinese LLM model” ranking based only on total parameters would put Kimi K3 first and end the discussion. That would not help a developer choose a model.

We considered five questions instead:

1. How does the model perform in independent evaluations rather than only its developer’s benchmarks?
2. Can it sustain repository-scale coding and repeated tool use?
3. How fast and expensive is it when accessed through an API?
4. Does it accept screenshots, diagrams or other visual input?
5. Can a company download, modify and commercially deploy its weights?

The last distinction matters. “Open,” “open weight,” and “open source” do not always grant the same rights. GLM-5.2 and DeepSeek V4 Pro use the permissive MIT license. MiniMax M3 publishes weights under its Community License. Qwen3.7 Max is proprietary. Kimi K3 now publishes its full weights under the custom Kimi K3 License, which permits broad use and modification but adds conditions for large Model-as-a-Service businesses and very large commercial products.

1. Kimi K3: Best Chinese LLM Overall

Moonshot AI built Kimi K3 for coding, knowledge work and reasoning tasks that may continue across many tool calls. Its headline specification is hard to miss: 2.8 trillion total parameters, a 1M-token context window and native understanding of text, images and video.

The architecture is more interesting than the raw size. Kimi K3 uses Kimi Delta Attention and Attention Residuals, activating 16 of its 896 experts for each token. That lets Moonshot scale the model without paying the full compute cost of a dense 2.8T-parameter model on every step. Moonshot’s Kimi K3 guide explains the architecture and context window.

Kimi K3 currently scores 57 on the Artificial Analysis Intelligence Index, ahead of the other four models in this list. It is also the only model here that combines that level of general reasoning with native visual input. That makes it particularly relevant for:

• Debugging a frontend from screenshots and runtime logs
• Rebuilding an interface from a visual reference
• Navigating a large repository over a long agent session
• Combining technical documents, diagrams and code
• Repeatedly testing and revising an application

There are three substantial trade-offs.

First, price. On GPT Proto, output costs $13.50 per million tokens. That is more than three times GLM-5.2's output rate and over fourteen times MiniMax M3's. A long agent can produce a lot of tokens while planning, editing, reviewing test results, and trying again.
Second, infrastructure. The full K3 weights are now public, but the repository is about 1.56 TB. Moonshot recommends supernode deployments with 64 or more accelerators. This is an open-weight model for infrastructure teams, not a laptop download.
Third, license. Kimi K3 is not MIT. Its custom license broadly allows use, modification, fine-tuning, and deployment, but adds conditions for large Model-as-a-Service businesses and very large commercial products. Legal review belongs in the deployment plan.

Verdict: Choose Kimi K3 when the task is difficult enough to justify the higher API bill or when access to a frontier open-weight checkpoint matters. Keep GLM-5.2 as the simpler open-weight default for routine coding agents.

2. GLM-5.2: Best Chinese Coding Model for Long-Running Agents

GLM-5.2 is the model I would start with for a serious coding agent.

Z.ai trained it specifically for long-horizon tasks rather than treating coding as another chat capability. It is a 753B-parameter Mixture-of-Experts model with about 40B parameters active per token, a 1M-token context window, native tool calling and MIT-licensed weights. Z.ai’s GLM-5.2 announcement describes its long-horizon positioning.

On the Artificial Analysis Intelligence Index, GLM-5.2 scores 51. More importantly for interactive development, independent measurements place its output speed near 200 tokens per second. That is considerably faster than MiniMax M3 and DeepSeek V4 Pro in the same dataset.

The combination suits work such as:

• Repository-wide refactoring
• Multi-file debugging
• Long terminal and test loops
• Migration work that spans code, configuration and documentation
• Commercial self-hosting or fine-tuning

The MIT license is a real advantage. A team can inspect the weights, host the model inside its own environment and adapt it without negotiating a proprietary per-seat agreement.

The cost is that GLM-5.2 is text-only. It cannot inspect a broken layout screenshot or reason directly over a design reference. Independent testing also finds it relatively verbose. Fast generation helps, but unnecessary output still increases the bill.

At $1.26 per million input tokens and $3.96 per million output tokens on GPT Proto, GLM-5.2 is not the cheapest model in this list. It is the most balanced open model for a coding agent.

Verdict: Use GLM-5.2 as the default when you need an agent to stay on task across a large repository. Switch elsewhere when visual input or the lowest possible token cost matters more.

3. Qwen3.7 Max: Best for Fast Hosted Coding Workloads

Qwen3.7 Max is Alibaba’s proprietary agent model. It provides a 1M-token context window, up to 65,536 output tokens and extended reasoning for coding and long-running execution. Unlike Qwen3.7 Plus, Max is text-only. Alibaba’s Qwen3.7 announcement positions Max as an agent foundation model.

Its most useful advantage is speed.

Artificial Analysis currently measures Qwen3.7 Max at about 204 output tokens per second, compared with roughly 65 for DeepSeek V4 Pro. It also scores 46 on the Intelligence Index, slightly ahead of both MiniMax M3 and DeepSeek V4 Pro.

The GPT Proto rate makes the model more interesting: $0.36 per million input tokens and $1.44 per million output tokens. That puts it close to MiniMax M3 on a realistic coding task while offering higher independent general-intelligence results.

Qwen3.7 Max fits:

• Fast code review
• Multi-file text-only refactoring
• Agent tasks where response time affects user experience
• High-volume hosted workloads
• Teams that want a strong API model without managing weights

The trade-off is control. Qwen3.7 Max is closed weight, so you cannot deploy it in a private cluster or fine-tune the base weights yourself. It also lacks Kimi K3 and MiniMax M3’s visual inputs.

There is also a freshness issue. Alibaba has already announced Qwen3.8 Max. Until that newer model has stable documentation, pricing and independent results, replacing a working Qwen3.7 integration would be premature.

Verdict: Qwen3.7 Max is the best choice here when hosted speed and low API cost matter more than self-hosting or multimodal input.

4. MiniMax M3: Best Value Chinese LLM for Multimodal Development

MiniMax M3 is easy to underestimate if you look only at its score of 44.

The model has approximately 428B total parameters and 23B active parameters, yet supports a 1M-token context and native text, image and video input. MiniMax Sparse Attention reduces attention compute at the full context length to roughly one-twentieth of the previous generation. MiniMax reports ninefold faster prefill and fifteenfold faster decoding than M2 at 1M context. The official MiniMax M3 model card contains the architecture and speed figures.

That makes MiniMax M3 a practical option for:

• Reading repository diagrams and UI screenshots alongside code
• Long-document and code analysis
• Multimodal bug triage
• High-volume coding subtasks
• Agents that need a long context without Kimi K3’s output price

On GPT Proto, the model costs $0.48 per million input tokens and $0.96 per million output tokens. It is the cheapest output model in this comparison.

The weights are downloadable, but there is a legal distinction worth preserving: MiniMax M3 uses the MiniMax Community License, not MIT. Open weights do not automatically mean unrestricted commercial use. A company planning to redistribute, fine-tune or self-host the model should review the license rather than relying on the word “open.”

On GPT Proto, text calls use the main MiniMax M3 endpoint. Image and file inputs are available through its dedicated image-to-text route under the same account.

Verdict: MiniMax M3 is the strongest cost-to-capability choice in the list. Pick it when you need multimodal input or expect to run a large number of agent subtasks.

5. DeepSeek V4 Pro: Best MIT Model for Backend Reasoning

DeepSeek is no longer the automatic answer to “What is the best Chinese coding model?” The competition moved.

DeepSeek V4 Pro is still a serious option. It is a 1.6T-parameter Mixture-of-Experts model with 49B active parameters, a 1M-token context window and MIT-licensed weights. It supports non-thinking, high-reasoning and maximum-reasoning modes through one model family. DeepSeek announced V4 Pro and V4 Flash in April 2026.

Its Artificial Analysis Intelligence Index score is 44, level with MiniMax M3 but below Qwen3.7 Max, GLM-5.2 and Kimi K3. Its output speed is also lower at about 65 tokens per second.

Why keep it in the top five?

DeepSeek V4 Pro remains well suited to text-heavy backend work:

• Algorithm and data-structure problems
• STEM reasoning
• Backend regressions
• Test generation and repair
• Private deployments that require a permissive license
• Output-heavy agent workflows where it is cheaper than GLM-5.2

On GPT Proto, it costs approximately $1.39 per million input tokens and $2.78 per million output tokens. GLM-5.2 has slightly cheaper input, but DeepSeek is cheaper on output.

Verdict: DeepSeek V4 Pro is no longer the broad winner, but it remains one of the better choices for backend reasoning when MIT-licensed deployment and predictable output cost matter.

One Real Coding Task, Five Different Choices

Consider a more realistic task than asking each model to build a to-do app:

A TypeScript SaaS application starts failing intermittently during its OAuth callback after an authentication-library upgrade. The coding agent must inspect the repository, trace the callback and session flow, identify the regression, change the implementation, add a failing test, run the suite and prepare a pull-request summary.

A useful initial instruction might look like this:

Investigate the OAuth callback regression in this repository.

Before editing:
1. Map the callback, session and token-refresh paths.
2. Identify the library upgrade that changed the behavior.
3. Reproduce the failure with a test.
4. Propose the smallest safe patch.

After editing:
1. Run the relevant unit and integration tests.
2. Report every file changed.
3. Explain remaining risks.
4. Do not claim success unless the tests pass.

The best model changes with the evidence and environment:

This is why I would not route every step to one model. A practical agent can use Qwen3.7 Max or MiniMax M3 for initial triage, escalate a difficult repository-wide fix to GLM-5.2, and reserve Kimi K3 for tasks that need visual reasoning.

What Would the Same Coding Task Cost?

Suppose the OAuth task consumes:

• 100,000 uncached input tokens
• 20,000 output tokens

Using the listed GPT Proto rates, the first-pass model cost would be:

For example, the Kimi K3 calculation is:

(0.1 × $2.70) + (0.02 × $13.50) = $0.54

This is not a claim that Qwen or MiniMax will solve the task for six cents. It is first-pass token arithmetic. If a cheaper model produces an incorrect patch and needs three retries, its real cost and completion time rise. Caching, reasoning modes, tool output and the number of agent turns also change the final bill.

The useful metric is not price per token. It is cost per accepted fix.

Why the Benchmark Winner May Not Be the Best Coding Agent

A coding model does not inspect files, execute commands or merge a patch by itself. The agent around the model controls:

• Which repository files enter the context
• How search results and terminal output are presented
• Whether edits use exact patches or rewritten files
• When tests are executed
• What happens after a failed command
• Whether the model can verify its own work
• How much context is preserved between turns

Put the same model inside two different agent systems and you can get very different results.

The 1M-token context shared by all five models creates another trap. One million tokens is capacity, not a recommendation to paste the entire repository into every request. Irrelevant generated files, dependency code and old logs can distract the model while increasing latency and cost.

A better coding agent retrieves relevant files first, keeps stable instructions near the beginning, summarizes old tool output and expands the context only when the investigation requires it.

One sentence version: choose the model and the agent together.

Which Chinese Coding Model Should You Use?

Choose Kimi K3 if you want the highest current overall capability, the task includes visual evidence or long reasoning, and you either prefer a hosted API or can support a 1.56 TB open-weight deployment. Review the custom license before building a commercial Model-as-a-Service product around it.

Choose GLM-5.2 if you are building a long-running coding agent and want the best balance of capability, speed, context, cost, and permissive deployment. Its MIT license and smaller footprint still make it my default self-hosted recommendation.

Choose Qwen3.7 Max if you want a fast, inexpensive hosted model and do not need downloadable weights or image input.

Choose MiniMax M3 if cost and multimodality matter. It is particularly attractive for high-volume development workflows that mix screenshots, files and code.

Choose DeepSeek V4 Pro if the work is backend-heavy and you want an MIT-licensed model that can run in your own infrastructure.

There is no permanent winner. Model releases are moving too quickly for that. What you can build is a routing layer that lets the winner change without forcing an application rewrite.

How to Call These Chinese LLM Models Through One API

GPT Proto exposes the five models through the same OpenAI-compatible chat endpoint. Here is a cURL request using GLM-5.2:

curl --location 'https://gptproto.com/v1/chat/completions' \
  --header 'Authorization: YOUR_GPTPROTO_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "glm-5.2",
    "messages": [
      {
        "role": "system",
        "content": "You are a senior software engineer. Separate confirmed evidence from hypotheses and never claim a test passed unless you saw its output."
      },
      {
        "role": "user",
        "content": "An OAuth callback began failing after an authentication-library upgrade. Return a JSON object with diagnosis_questions, files_to_inspect, likely_failure_modes, and test_plan."
      }
    ],
    "stream": false
  }'

To try another model, change the model value:

Capitalization matters for MiniMax-M3.

You can compare the wider GPT Proto model collection or check the current pay-as-you-go pricing before routing production traffic.