Hooking SJTU’s Free On-Campus LLM API into TRAE for Surprisingly Smooth AI Coding

SJTU has never been short on usable resources, and its local large-model platform is one of those things that turns out to be genuinely practical once you start playing with it.

I’m not an AI or LLM specialist by any stretch. I just fall into the classic category of "not that good, but very curious," so I keep tinkering with odd little setups. I actually wired the university’s local models into my IDE quite a while ago, but between being busy and still stumbling through configuration pitfalls, I never got around to writing it down. Now that there’s finally a bit of breathing room, it’s worth documenting—otherwise it would probably stay forever in the "I’ll write this later" pile.

A lot of code polishing work I did in previous projects was already assisted by the school’s locally deployed models. They didn’t handle everything perfectly, of course—some pieces still needed manual fixing—but as a productivity tool inside an IDE, the effect is obvious, especially if you’re not from a hardcore computer science background.

What this platform actually is

The models are provided through SJTU’s locally deployed "Zhiyuan No.1" platform, part of the university’s "Jiao Wo Suan" computing center.

According to the university’s own introduction, "Zhiyuan No.1" is a new core component of the campus-level intelligent computing system and currently the largest domestic AI computing infrastructure among Chinese universities. Construction started in December 2024 under a build-and-use-in-parallel approach. In February 2025, it launched what was described as the first fully local deployment of DeepSeek at a domestic university, and it has since supported a number of high-level research projects. The platform has now completed local deployment of eight mainstream large models, including DeepSeek, and opened them to faculty and students. In 2025, it reportedly served 38,000 users on campus and became a key computing base for the university.

For IDE integration, the same basic idea can also be used with tools such as Claude Desktop or Claude Code. But for most non-CS users, TRAE currently feels like one of the easier AI IDEs to get along with. I used Cursor before, but it has become increasingly restrictive—for example, some custom model address features now sit behind the Pro tier.

Access rules before you start

There are a couple of limits worth knowing up front:

• The API service is only accessible on the campus network, or off campus through VPN.
• Selling access, leaking the api-key, or abusing the computing resources is explicitly prohibited. Violations can lead to cancellation of the application and future extension eligibility, along with further accountability.

There is also a rate limit. Earlier it was 100 requests per minute, but later it was reduced to 10 requests per minute. That’s a lot tighter than before, though it may simply reflect the pressure of running larger models locally.

Models currently available

The platform currently exposes the following models:

<table> <thead> <tr> <th>Model</th> <th>Parameters</th> <th>API name</th> <th>Mode</th> <th>Strengths</th> <th>Context length</th> </tr> </thead> <tbody> <tr> <td>DeepSeek V3.2 (standard)</td> <td>685B</td> <td>deepseek-chat</td> <td>Non-reasoning mode</td> <td>General text processing</td> <td>32k</td> </tr> <tr> <td>DeepSeek V3.2 (reasoning)</td> <td>685B</td> <td>deepseek-reasoner</td> <td>Deep reasoning mode</td> <td>Complex logic and in-depth inference</td> <td>32k</td> </tr> <tr> <td>MiniMax-M2.7</td> <td>230B</td> <td>minimax or minimax-m2.7</td> <td>Text generation</td> <td>Agent-style tasks</td> <td>192k</td> </tr> <tr> <td>GLM-5.1</td> <td>754B</td> <td>glm or glm-5.1</td> <td>Text generation</td> <td>Coding ability and long-horizon tasks</td> <td>128k</td> </tr> <tr> <td>Qwen3.5-27B</td> <td>27B</td> <td>qwen or qwen3.5-27b</td> <td>Multimodal</td> <td>Vision and text understanding</td> <td>256k</td> </tr> </tbody> </table>

In practice, I added all of them. If one model doesn’t behave well on a task, another one sometimes does. No need to be overly loyal to a single option.

How to apply for the API

Open the web version or mobile app of the campus service platform and search for “Zhiyuan No.1 AI Model API Application (Test)”. Then submit the workflow request for model API access.

Fill in the form and submit it. Once approved, you’ll receive the base_url and api-key through the email address you provided and via the campus platform’s internal notifications.

The notification email will look roughly like this:

Installing TRAE or TRAE SOLO

My personal recommendation is to use the standard version of TRAE.

SOLO feels more oriented toward pure vibing-style coding, while the regular version is easier to accept if you actually want to inspect code, work like a developer, or keep some control over what’s happening.

I’ve generally been pretty willing to use AI tools because I don’t romanticize them. Their so-called intelligence is still just a giant statistical engine predicting the next token. Once you accept that, it becomes much easier to treat them for what they are: efficiency tools.

Installation itself is nothing special—download it from the official site and click through the usual setup steps unless you want to change the installation path.

Configuring the campus models in TRAE

This is the useful part. TRAE can talk to these models because the campus API follows the OpenAI-compatible standard.

Open the model settings and add a new model:

Then configure it like this:

• Provider: OpenAI
• Model: Custom model
• Model ID: deepseek-chat or deepseek-reasoner or minimax or glm or qwen
• API key: paste the value from the email
• Custom request URL: https://models.sjtu.edu.cn/api/v1/chat/completions

I added them one by one, all in the same way.

Once that’s done, you can start using AI assistance directly in the editor.

A few practical impressions

I’ve already loaded several of the school-provided models into the IDE. Sometimes one is unhelpful and another one immediately gets the job done. It’s really a matter of trying more than one.

Personally, I still think tools like CodeX and Claude are stronger in some scenarios, and Gemini is also usable—but those are foreign closed-source models, which is a separate discussion. For a campus-provided, free, locally deployed option that can be plugged straight into an IDE, this setup is already pretty capable.

If your goal is to have AI help read, modify, and generate code inside the development workflow, this is one of the most practical ways to do it with the resources already available on campus.