Agent Frameworks for Chinese LLM APIs (2026)

Every major open-source agent framework worth naming in 2026 is model-agnostic, so CrewAI, LlamaIndex, SWE-agent, and Pydantic AI all accept Chinese LLM APIs like DeepSeek, Qwen, GLM, Kimi, and MiniMax through OpenAI-compatible endpoints. The compatibility question people ask has a boring answer: they all work. The interesting part is why, and which backing model costs the least once you plug it in.

An open-source agent framework is developer tooling that orchestrates LLM calls, tool use, and multi-step reasoning while staying independent of any single model provider. That independence is the whole reason a framework can point at a Chinese API at all. None of the four below hard-code a vendor; each takes a base URL and an API key, and a Chinese provider that speaks the OpenAI chat-completions schema slots in where GPT-4o would have gone.

Agent frameworks for Chinese LLMs (verified 2026-07)

We pulled the four most-cited open-source agent frameworks and checked each repository directly. The pattern is consistent: all four are Python, all four are MIT-licensed, and none of them care which model sits behind the endpoint.

The MIT license across the board matters more than it looks. It means you can run any of these against a Chinese API in a commercial product without a copyleft obligation reaching back into your own code. That removes the second question buyers usually have after "does it connect."

We checked the repositories on 2026-07-10 (first-hand evidence)

Documentation ages fast, so instead of quoting a changelog we read the live GitHub state on 2026-07-10. As of that date, per the GitHub API, CrewAI showed 55,260 stars, an MIT license, and a latest release of 1.15.2 dated 2026-07-08. Pydantic AI was moving even faster: its v2.8.0 tag shipped on 2026-07-10, the same day we checked. Star counts and release dates move, so treat these as a dated snapshot rather than a standing figure.

According to the CrewAI repository, the project is MIT-licensed and written in Python, which is what lets it wrap a Chinese chat-completions endpoint without a license conflict. The recent 1.14.x line added pluggable model backends and a Chat API, and a pluggable backend is exactly the seam a DeepSeek or Qwen base URL drops into.

According to the LlamaIndex repository, the framework carries an MIT license and reached v0.14.23 on 2026-06-24. Its Workflows layer, which went to 1.0 in mid-2026, models multi-agent flows as event-driven steps, and each step's LLM call is provider-agnostic. That is the design choice that keeps a Chinese model swap to a one-line configuration change rather than a rewrite.

SWE-agent is the outlier in release cadence. Its latest tagged release, v1.1.0, dates to 2025-05-22, older than the others by more than a year, yet the repository was still being pushed to as of 2026-07-07. A quiet release tag is not a dead project; it usually means the maintainers ship from the main branch and tag infrequently. SWE-agent was built by Princeton researchers as a coding agent with an Agent-Computer Interface, and that interface is what it uses to read and edit real repositories.

Why OpenAI-compatibility is the whole game

The reason this cluster of frameworks "just works" with Chinese models is that DeepSeek, Alibaba's Qwen, Zhipu's GLM, and the others publish an endpoint that mirrors the OpenAI chat-completions request and response shape. A framework that already speaks that dialect does not need a Chinese-specific adapter. You change two values, the base URL and the key, and the SWE-agent loop or the CrewAI crew keeps running unchanged.

This is also why we do not frame any of these as a China-only tool. They are general agent frameworks that happen to accept Chinese endpoints as one option among many. The neutrality cuts both ways: the same LlamaIndex workflow can call GLM-5 today and Claude tomorrow with a config edit.

Pydantic AI is worth a separate note because it comes at the problem from types. Pydantic AI is a type-safe agent framework that validates model inputs and outputs against Python type hints, which reduces the class of runtime errors that show up when you swap one model's quirks for another's. If you are moving a workload from a Western model to a Chinese one and worried about output-shape drift, that validation layer is a practical safety net.

Which backing model costs the least

Compatibility is free; tokens are not. Once any of these four frameworks is pointed at a Chinese endpoint, the bill is set by the model you chose, not the framework. According to DeepSeek's API documentation, DeepSeek V4-Flash is priced at $0.14 per million input tokens and $0.28 per million output tokens, which makes it the cheapest first-party backing model in the mainstream Chinese lineup for a general agent loop.

For context, GPT-4o runs $2.50 input and $10 output per million tokens, so a V4-Flash-backed agent pays roughly one-eighteenth the input rate for the same orchestration code. That gap is the part most teams underestimate when they benchmark frameworks in isolation and forget the model underneath is where the money goes. If your agent is coding-heavy, a coding-tuned Chinese model changes the arithmetic again, and the DeepSeek API pricing hub breaks down where those rates land per workload.

Choosing among the four

Pick by job, not by star count. Choose CrewAI if you want role-based multi-agent crews with the largest community and the most recent release cadence. Choose LlamaIndex if your agent is retrieval-heavy and you want event-driven workflows. Choose SWE-agent if the task is autonomous code repair on real repositories. Choose Pydantic AI if type-safety and validated outputs matter more than breadth. All four accept the same Chinese endpoints, so the framework decision and the model decision are genuinely separate, and you can make them one at a time.

FAQ

Do open-source agent frameworks work with Chinese LLM APIs? Yes. CrewAI, LlamaIndex, SWE-agent, and Pydantic AI are all model-agnostic and accept any OpenAI-compatible endpoint, so DeepSeek, Qwen, GLM, Kimi, and MiniMax plug in by changing only the base URL and API key.

Which agent framework has the most GitHub stars in 2026? Of these four, CrewAI leads. We checked on 2026-07-10 and, per the GitHub API, it showed 55,260 stars, ahead of LlamaIndex at 50,751, SWE-agent at 19,757, and Pydantic AI at 18,300. Counts move over time.

Are these frameworks free to use commercially with a Chinese model? Yes. All four are MIT-licensed, a permissive license that allows commercial use without a copyleft obligation, so running them against a Chinese API in a paid product carries no license conflict.

What is the cheapest Chinese model to run an agent on? Among first-party endpoints, DeepSeek V4-Flash is the cheapest general option at $0.14 input and $0.28 output per million tokens, roughly one-eighteenth of GPT-4o's input rate for the same framework code.

Do I need a special adapter for Chinese models? No. Because the major Chinese providers expose an OpenAI-compatible chat-completions schema, any framework that already targets OpenAI works with no China-specific adapter. You reuse the existing client and only change the endpoint URL.

This is part of the open-source AI agents hub, and for a broader provider comparison see the best Chinese LLM API guide.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in the Chinese LLM ecosystem and AI infrastructure pricing. Last verified: 2026-07-10.