AI provider

The assistant in PKU harness needs a large language model to understand you and to act. The app doesn't include a model, and we don't run one. You open an account with a model provider, create an API key there, and enter it in the app. An API key is a long string, often starting with sk-, that works like a password for the provider's service.

The key is stored in the phone's secure storage and only sent to the provider you choose. Your chat messages go straight from your phone to that provider. Providers charge for use: you top up an account and each reply costs a small amount.

Recommended providersCopy Anchor Link

Any service with an OpenAI-compatible API works. For students in mainland China, these three are easy to pay for and reachable without a VPN.

Model names change often. Check the provider's model list for the current ones: Bailian models, Qianfan models.

Which one to chooseCopy Anchor Link

DeepSeek is the simplest choice, and the one we recommend first. The app knows its models, so:

• You don't need to type a base URL or a model name.
• The app knows each model's context size (1M tokens) and whether it reads images. DeepSeek-V4.1-Flash accepts images, so you can send photos and Copilot can use screenshots. DeepSeek-V4-Pro doesn't.
• Settings → Usage estimates the cost from DeepSeek's published prices, and Settings → AI provider shows your account balance.

Alibaba Cloud Bailian and Baidu Qianfan work through the OpenAI-compatible provider. Choose one of them if you already have an account there or prefer their models. Bailian has one more advantage: the same account gives you a key for voice input. The app knows less about models connected this way:

• It assumes a context size of 128k tokens, and compacts the chat based on that.
• It doesn't offer image attachments for these models, so you can't send photos and Copilot can't use screenshots with them.
• Settings → Usage counts their tokens but can't estimate the cost, and the app can't show the balance. Check both in the provider's console.

Get a DeepSeek API keyCopy Anchor Link

Get an Alibaba Cloud Bailian API keyCopy Anchor Link

Bailian bills each call to your Alibaba Cloud account. New accounts sometimes get a free quota for some models; the console shows what you have.

The base URL above is the Beijing address from Bailian's API key guide. Bailian's OpenAI compatibility page also lists a workspace address, https://<WorkspaceId>.cn-beijing.maas.aliyuncs.com/compatible-mode/v1, which it recommends for stability. Either works.

Get a Baidu Qianfan API keyCopy Anchor Link

The base URL and key format come from Qianfan's quick start. Calls are billed to your Baidu AI Cloud account.

Enter the key in the appCopy Anchor Link

You can connect a provider during first-run setup or later in Settings. The Settings screen has every option.

1. Open Settings → AI provider.
2. Tap a provider: DeepSeek or OpenAI-compatible.
3. Under Model, tap a model, or type its name in Model ID. For OpenAI-compatible, type the name; once the test succeeds, the app also lists the models the provider reports.
4. For OpenAI-compatible, fill in Base URL. For DeepSeek, the address is fixed and shown as "Endpoint: https://api.deepseek.com".
5. Paste the key into API Key. A saved key is marked "Saved in secure storage".
6. Tap Save and test. The app sends a very short test message and, on success, shows "Connected" with the model's reply. To save without testing, tap Save only.

If the test fails, the message says why. See Troubleshooting → AI provider.

The assistant can't read or enter API keys. If you ask it to change your key in chat, it opens this screen for you to do it.

Account balanceCopy Anchor Link

With DeepSeek selected and a key saved, Account balance appears on the AI provider screen. It shows the balance and how much of it was granted and topped up ("Granted … · Topped up …"). Tap it to refresh. If the provider says the balance can't cover requests, the screen says so; top up on the provider's website to keep chatting.

For other providers, check the balance in their console.

Switch modelsCopy Anchor Link

On the Chat tab, tap the model label below the message box, or type /model, to choose another model. The new model is used in this chat from your next message, and becomes the default for new chats. See Chat → Model and context.

You can also ask the assistant, for example "Switch to DeepSeek-V4-Pro". It changes the setting and tells you what it changed; see Settings → Changing settings in chat.

Context budgetCopy Anchor Link

The model sees your chat as its context. Context budget at the bottom of Settings → AI provider limits how much of it the app sends. The options are 32k, 64k, 128k, 256k and Model maximum. The default is Model maximum.

With Model maximum, the app uses the model's whole context and summarizes older messages when the chat reaches 90% of it. On DeepSeek's 1M context that is about 900k tokens. A smaller budget means faster and cheaper replies, but the assistant sees less of the chat directly. Summarizing also happens in the middle of a long task, so the task carries on.

Usage and costCopy Anchor Link

You pay the provider for what you use, by the number of tokens (pieces of text) sent and received. Longer chats, attachments and long Copilot tasks use more tokens.

Settings → Usage shows token usage, active hours and top tools, for 7 days, 30 days or All time. For models with published prices, such as DeepSeek's, it estimates the cost. The estimate is only a guide; your provider's bill is what counts. See Settings and your data → Usage.