Practical Guide to Free Chinese LLM APIs: From Registration to Side-by-Side Model Comparison

Why Compare Multiple Models?

With the rapid proliferation of Chinese open-source large language models, developers and enterprise users face a practical dilemma: different models excel in different dimensions—code generation, logical reasoning, or Chinese language understanding. Choosing a model based on benchmark leaderboards or marketing claims often diverges significantly from real-world experience. The most reliable approach is to send the same set of test prompts to multiple models and compare their outputs under realistic conditions.

Traditionally, such model comparison experiments required registering separately with each vendor’s platform, obtaining individual API keys, and integrating with different API specifications—a tedious and costly process. With NVIDIA NIM (NVIDIA Inference Microservices) now hosting several Chinese open-source models, this picture has changed: users can obtain a single API key, call multiple mainstream Chinese models through a unified interface, and view their answers side-by-side in a multi-model chat client, dramatically improving evaluation efficiency.

Platform and Tool Overview

NIM Platform: A Supermarket-Style Entry Point for LLMs

NIM is NVIDIA’s model inference service platform—essentially a unified, packaged model marketplace. It deploys open-source models from different vendors and architectures on a common backend inference environment and exposes an OpenAI-compatible API. For users, this means one set of credentials is enough to call all hosted models, eliminating the need to adapt to each vendor’s distinct API format.

The platform currently hosts leading Chinese open-source models such as Zhipu GLM-5 and Moonshot Kimi K2.5, and offers free usage quotas. For individual developers, academic researchers, and small teams, this provides an extremely low-cost channel for evaluation and experimentation.

Client Choice: Cherry Studio

An API alone is not enough—a graphical client is needed for day-to-day conversations. Cherry Studio is a desktop application that supports multiple model service providers. Its core value lies in side-by-side multi-model comparison: one question can be dispatched simultaneously to several models, with responses arranged as horizontal cards so differences are immediately visible. This interaction model is naturally suited to model evaluation scenarios, eliminating the need to switch windows and compare outputs manually.

Complete Setup Guide

Step 1: Register a Platform Account

Open the NIM platform website (build.nvidia.com) and click Login in the top-right corner. Enter your email, set a password, and complete the basic registration. You will receive an email containing a 6-digit verification code; enter it to activate your account.

Important note: After registration, a “Verify” entry appears in the top-right corner. You must click it and complete phone number verification. International numbers are accepted, and verification SMS messages are delivered without issue. Skipping this step will prevent successful API key generation or limit API access.

Step 2: Generate an API Key

After logging in, click your avatar in the top-right corner and navigate to the “API Keys” management page. Click the “Generate API Key” button, assign a recognizable name to the key, and set the expiration to “Never” so it remains valid long-term. The generated key will be a string beginning with nvapi-.

This key is your sole credential for accessing all free models on the platform. Copy and store it securely right away. Once you close the page or refresh, the plaintext key will no longer be displayed, and you will need to generate a new one.

Step 3: Configure the Client

1. Download and install Cherry Studio (from cherry-ai.com).
2. Open the application, go to “Settings” → “Model Services” in the bottom-left corner, and select “NVIDIA” as the service provider.
3. Fill in the following parameters:
   • API Key: paste the nvapi-xxx string obtained in Step 2.
   • API Address: https://integrate.api.nvidia.com
   • Model IDs: fill in as needed. Recommended models:
     • z-ai/glm5 — Zhipu’s latest flagship
     • moonshotai/kimi-k2.5 — Moonshot’s latest model
     • minimaxai/minimax-m2.1 — MiniMax’s m2 series (note: NIM currently hosts m2.1; m2.5 has not yet been added—check the platform for updates)
4. After saving the configuration, you can select the configured models in the chat interface and start conversations.

Side-by-Side Multi-Model Comparison in Practice

Cherry Studio’s comparison feature is a powerful evaluation tool. At the bottom of the chat input box, you can select multiple configured models. When you hit send, the same question is distributed in parallel to all selected models. The answers appear as horizontal cards—one card per model—so you can directly compare output quality on the same question without switching windows.

This mode is especially well suited for the following evaluation tasks:

• Code generation: Have multiple models write a function for the same task, then compare coding style, error handling, and comment quality.
• Logical reasoning: Ask a question requiring multi-step deduction and observe whether each model’s reasoning chain is complete and whether the conclusion is reliable.
• Chinese-language context understanding: Use Chinese proverbs, puns, or industry jargon to gauge each model’s grasp of native language culture.
• Long-document summarization: Feed the same technical document to all models and compare the structural quality and key-information retention rate of each summary.

Notes and Practical Recommendations

1. Free quota and rate limits: NIM’s free tier is sufficient for personal experimentation. For high-concurrency or production use, be aware of platform rate-limiting policies. Occasional response delays may occur during peak hours, but routine evaluation is unaffected.
2. Model IDs change over time: The platform continuously adds new model versions (such as MiniMax M2.5). Check the model list page periodically and update your client configuration accordingly.
3. Think in terms of multi-model orchestration: Do not commit to a single model for everything. Different models have different strengths. In real projects, design a routing strategy: use model A for code generation, model B for document understanding, and take advantage of each model’s specialty.
4. Control variables for rigorous evaluation: For a fair model comparison, ensure sampling parameters such as temperature and top_p are identical across models (Cherry Studio allows these adjustments in settings) to keep the comparison equitable.

Conclusion

By combining NIM’s aggregation capability with Cherry Studio’s side-by-side multi-model interaction, developers can build a low-cost Chinese LLM evaluation workflow. This is not just a model-selection tool—it is also a window into understanding the architectural differences among major Chinese LLMs. When two models’ answers appear side by side on the same screen, differences in training data, alignment strategy, and reasoning habits become evident in the concrete output. It is recommended that readers set up the environment following this guide and run a comparison round with their own real business questions. First-hand experience will always be more convincing than reading evaluation reports.