Model Selection and Customization: 14 Built-in Models, API Key, and Ollama
The execution mode was settled in the previous lesson; this page answers "which model to send on the job." For the same task, sending a Fast-tier or Ultimate-tier model, using built-in DeepSeek or local Ollama, can make a huge difference in speed, quality, and the credit bill. WorkBuddy Models come in two layers: built-in models—14 of them—work out of the box and charge credits by multiplier; custom models are ones you connect yourself—API Key, Tencent Cloud plans, or local Ollama all work—and they don't consume WorkBuddy credits. In 12 minutes, get a clear handle on both "choosing a model" and "connecting a model" at once.
This is lesson 5 of "Daily Use," based on WorkBuddy v5.9.3, with operations based on the official "Model Configuration" documentation and the App / mini-program "Choose a Model" documentation.
- Still unclear about execution modes? Go back to 2.4 Three Execution Modes: Default (Agent), Plan, and Ask Only
- Want an overview of this stage first? Go back to 2.0 Daily Use: Turn WorkBuddy Into Your Office Sidekick
What You'll Be Able to Do After This Lesson
- Understand the two layers of WorkBuddy models: 14 built-in models that work out of the box, and custom models connected as needed
- Switch between Fast / Balanced / Ultimate modes in the model selector and pick the right built-in model by task (checking capability tags, multipliers, and busy status)
- Connect custom models via provider presets + API Key, or local Ollama, and manage the keys stored in the local models.json
- Explain clearly how thinking intensity, context window, and Economy mode each affect credit consumption
- Fix 4100 errors or grayed-out models on your own when they occur
Your Current Pain Points
- The task isn't hard, but you keep using the same model, burning plenty of credits even on simple Q&A
- You want to use a model plan purchased company-wide, or keep content entirely on your machine, and don't know whether WorkBuddy can connect to it
- You added a custom model, but it's grayed out and unselectable, and a 4100 pops up in the conversation
When to Use This Technique
When you need to:
- Match different models to different tasks—simple polishing should be fast and cheap, complex analysis should be stable
- Connect to third-party model services: API Key, Tencent Cloud Token Plan, Coding Plan
- Keep code and conversations on your machine (local Ollama), usable in intranet and offline environments
And you don't want to:
- Hand-edit configuration files or dig through a pile of parameters just to use a certain model
- Hand your API Key over to a cloud you don't trust
🎒 Preparation Before Starting
- Completed 2.0 Daily Use: Turn WorkBuddy Into Your Office Sidekick; ideally you've also studied 2.4 Three Execution Modes
- Installed WorkBuddy on your computer and logged in, able to create tasks and find the input bar
- (Optional) To connect a custom model: have the API Key from your provider ready; to connect Ollama: Ollama is installed and running on your machine
- 12 minutes of reading time
💡 Does it cost credits?
Built-in models are billed by credits, with different multipliers per model (x0.06~x1.62); custom models don't consume WorkBuddy credits—all costs incurred (tokens, subscriptions, etc.) are paid by you directly to the third party (per the official statement). See 2.6 Credits and Pricing for credit rules.
Core Idea
Mental model: WorkBuddy models = a selector in the conversation + a configuration page in settings.
| Layer | What it is | How to use it |
|---|---|---|
| Built-in models | The 14 models already connected by the official side (Hunyuan, GLM, Kimi, MiniMax, DeepSeek) | Pick directly in the model selector, billed by credit multiplier |
| Custom models | Third-party / local models you connect yourself | Configure in the "Models" section of the settings page; costs are paid to the third party and don't consume WorkBuddy credits |
The selector also has three mode tiers (Fast / Balanced / Ultimate), which decide which capability tier of model gets dispatched for the job. The official advice is blunt: when unsure which tier to pick, use Balanced.
Let's clear up three easily confused "modes" all at once:
| Name | What it governs | Where |
|---|---|---|
| Execution mode (Agent / Plan / Ask) | How the task proceeds | Left of the input box + menu (previous lesson) |
| Model mode (Fast / Balanced / Ultimate) | Which tier of model to dispatch | Model selection panel (this lesson) |
| Thinking intensity / context window | How deeply a single model thinks and how long it remembers | Model parameter configuration (this lesson) |
Security baseline (official statement): configuration parameters for custom models (including API Keys) are stored only locally in ~/.codebuddy/models.json and not uploaded to the cloud; when in use, WorkBuddy acts only as a communication link, forwarding your input directly to the model you configured, and the output is returned directly by that model.
Follow along
Step 1: Open the WorkBuddy model selector and pick a mode tier (Fast / Balanced / Ultimate)
Why: Different tasks vary greatly in what they demand from a model — pick the tier first, then the specific model, so you don't waste credits.
Steps:
- Create a new task and click the model name in the input bar (e.g., "Auto ▾") to open the model selection panel
- In the panel, first look at the three mode tiers (official comparison table):
| Mode | Applicable scenarios | Characteristics |
|---|---|---|
| Fast mode | Simple, well-defined tasks: everyday Q&A, format conversion, small code edits, copy polishing | Fastest responses, lowest consumption |
| Balanced mode | The vast majority of tasks: writing docs, analyzing data, multi-file code changes | Balances quality and speed; use it when unsure which tier to pick |
| Ultimate mode | Complex tasks: cross-module refactoring, multi-step reasoning, analyses requiring repeated verification | Slightly slower responses, higher consumption, but more stable results on complex tasks |
- Unsure which tier to pick: Balanced. For light work like everyday Q&A and polishing, you can switch to Fast to save credits; for hard jobs with multiple steps and repeated verification, go with Ultimate
You should see: After opening the panel, you can see the three mode tiers and the model list; after switching modes, subsequent conversations run at the new tier.
- I opened the model selection panel and saw the Fast / Balanced / Ultimate tiers
Step 2: Pick a built-in model based on the task (capability tags, multipliers, and busyness)
Why: The 14 built-in models each have their strengths, with multipliers ranging from as low as x0.06 to as high as x1.62 — picking the right one saves money and gets the job done.
Steps:
- Check capability tags: "Reasoning" = supports logical reasoning, task planning, and multi-step execution; "Vision" = supports image understanding (pick these for photo and screenshot scenarios)
- Check credit multipliers: the lower the number, the lower the consumption; it's shown below the model name (reference multipliers from the official mini-program docs are in the table below; on desktop, go by what the panel actually displays)
- Check busyness: hover over a model and the info card will show its "busyness"
List of the 14 built-in models (official desktop docs, v5.9.3):
| Model | Description | Available features |
|---|---|---|
| Hy4 preview | Hunyuan's new-generation model | Image input, thinking mode |
| Hy3 | Hunyuan thinking model with enhanced reasoning capabilities | Image input, thinking mode |
| GLM-5.3 | GLM New-generation model supporting a 1M context window | Image input, thinking mode |
| GLM-5.3-Flash | GLM Fast version of the new-generation model | Thinking mode |
| GLM-5.2 | 1M context, excels at long-horizon tasks | Image input, thinking mode |
| GLM-5.1 | Balanced capabilities, suitable for everyday use | Reasoning model |
| GLM-5v-Turbo | Natively multimodal model | Image input, reasoning model |
| MiniMax-M3 | Natively multimodal, excels at coding and agent tasks | Image input, reasoning model |
| MiniMax-m2.7 | Balanced capabilities, suitable for everyday use, great value for money | Image input, reasoning model |
| Kimi-K3 | Excels at complex, long-horizon autonomous tasks, with outstanding front-end development skills and excellent performance in knowledge work and research reasoning | Image input, reasoning model |
| Kimi-K2.7-Code | Multimodal model optimized for programming scenarios | Image input, reasoning model |
| Kimi-K2.6 | Multimodal model, suitable for everyday tasks | Image input, reasoning model |
| Deepseek-V4-Flash | DeepSeek flagship model, supports a 1M context window, speed-first | Thinking mode |
| Deepseek-V4-Pro | DeepSeek flagship model, supports a 1M context window | Thinking mode |
Credit multiplier reference (official mini-program docs; for newer models like Hy4 preview and the GLM-5.3 series, go by what the panel actually displays):
| Model | Credit multiplier |
|---|---|
| Deepseek-V4-Flash | x0.06 |
| Deepseek-V4-Pro | x0.16 |
| MiniMax-M3 | x0.25 |
| Kimi-K2.6 | x0.52 |
| Kimi-K2.7-Code | x0.57 |
| GLM-5.2 / GLM-5.1 | x0.79 |
| GLM-5v-Turbo | x0.95 |
| Kimi-K3 | x1.62 |
| Hy3 | free for a limited time |
| Auto / MiniMax-m2.7 | The mini-program documentation does not label the rate multiplier; go by what the panel actually displays |
Four busyness states (official):
| State | Description |
|---|---|
| Idle (green) | Normal load |
| Slightly congested (orange) | Load is high; responses may slow down |
| Extremely congested (red) | Load is very high; switching to another model is recommended |
| Unavailable | The model service is abnormal and temporarily unavailable |
Extremely congested models also have a red icon next to their name in the list, visible without hovering. Busyness is for reference only and does not prevent you from selecting a model; it is not displayed when load information cannot be retrieved.
Official money-saving tips (App documentation): for everyday simple tasks, pick a cost-effective reasoning model, and switch to a "Vision" model only when image understanding is needed; if you're not sure which to choose, use the default Auto (smart automatic selection), and the system will automatically match the optimal model.
💡 Want to use DeepSeek? It's built in
No need to find an API Key yourself—Deepseek-V4-Flash (x0.06) and Deepseek-V4-Pro (x0.16) in the built-in list work out of the box; the former prioritizes speed, and both support a 1M context window. To connect to other DeepSeek endpoints, use the custom model setup below.
💡 Enterprise users: can't select a model?
Models disabled by the enterprise administrator in the backend are simultaneously hidden or grayed out in the client's model selector (including auto). If you find a model cannot be selected, first check with your enterprise administrator to confirm whether that model is enabled.
You should see: when you hover over a model, an info card appears showing the busyness level; the rate multiplier is labeled below the model name.
- I can pick a built-in model for a task based on "capability tags + rate multiplier"
Step 3: Fine-tune thinking effort and context window as needed
Why: these two parameters determine how deeply the model "thinks" and how long it "remembers," and directly determine the credit consumption per turn—the defaults suit the vast majority of tasks, so understand them before touching them.
Thinking effort (official wording):
- Meaning: the reasoning investment the model makes before giving an answer; the selectable levels depend on the tiers the model itself supports
- Higher effort → more analysis and self-checking, suited to tasks with long logic chains that are prone to error; lower effort → tends to answer directly, with faster responses
- A level suited to the vast majority of tasks is selected by default, and separate adjustment is generally unnecessary
- Mobile App entry: configuration entry in the top-right corner of the model selection panel → Thinking effort; the tiers depend on the model (e.g., Low / Medium / High / Ultra)
Context window (official wording):
- Each turn of the conversation sends the full current context to the model—the larger the window, the more credits a single turn consumes
- Official reminder: bigger is not always better—a larger window doesn't make the model smarter, it only multiplies the per-turn consumption
- Multi-tier models: click the context usage ring in the input box, and switch tiers at the bottom of the "Context Usage" panel (e.g., Default / Moderate / Ultra-long)
- Rules: hover over a tier to see its description; changes are saved immediately; while a conversation is in progress, the new tier takes effect from the next turn; a tier applies only to the current task, and reverts to the model's default tier after switching models or creating a new task
- Models with only one tier do not provide a switching entry
- Recommendation: keep the default tier; it's sufficient for the vast majority of scenarios; only increase it when processing ultra-long documents or when the model needs to fully remember a large amount of historical content
Economy mode (a handy feature to know):
- The "Max mode" toggle in the original model selection panel has been moved to Settings → Agent → Economy mode (enabled by default for personal edition)
- When enabled: the system automatically compresses the context at appropriate moments, reducing repeatedly transmitted content in long conversations and lowering credit consumption
- When disabled: the full context is retained throughout with no compression, so no details are lost, but credit consumption for long conversations increases noticeably
- Recommendation: keep it enabled; version upgrades will not override your existing choice
You should see: clicking the context usage ring in the input box opens the "Context Usage" panel with tier switching at the bottom.
- I found the context usage ring and know that tiers apply only to the current task and revert to default after switching models
Step 4: Add a custom model: provider presets + API Key
Why: to use models outside the built-in list (e.g., a package your company purchased centrally), you don't need to touch configuration files—just a few clicks in the settings page's graphical interface; the official wording is "no need to manually edit configuration files."
Steps:
- Open the "Models" section of the settings page (the model selector in the conversation interface also has a quick entry to jump to the configuration page)
- Add a model, selecting from the provider list: the URL, model list, and capability tags (tool calling, image input, reasoning mode, etc.) are filled in automatically, and you only need to fill in the API Key to save and use it
- Choose an access method based on the service you have:
| Access method | Who it's for | What to fill in |
|---|---|---|
| Tencent Cloud Token Plan | Those who have purchased a Tencent Cloud Token Plan package (built specifically for AI coding and WorkBuddy scenarios, compatible with mainstream LLMs such as Hunyuan, MiniMax, Kimi, and GLM) | First select the package type (Enterprise Professional / Enterprise Lite / General Personal / Hy Personal), then fill in that package's API Key, and finally select the models the package supports (available models differ between packages) |
| Coding Plan | Those who have purchased a dedicated AI Coding subscription (supports Coding Plans from Tencent Cloud, Zhipu, and Kimi) | The corresponding package's credentials and models |
| Custom API | Those using APIs from other major vendors (built-in preset entries for major vendors, one-click access) | API Key |
| Custom / Custom | Services not in any preset list | Manually fill in the URL, API Key, and model name |
- Save. The configuration is persisted automatically; when you select a standard provider, capability tags such as tool calling and image input are written automatically, with no manual configuration needed
Security tips (key points of the official wording):
- Do not share your API Key with others—it is tied to your account and package quota; leaking it may allow others to steal your quota and incur extra costs or data security risks
- Confirm the provider is trustworthy—before connecting, read the model service provider's terms of service and privacy policy, and confirm they meet your requirements for data security and compliance
💡 Where is the configuration stored? Does it go to the cloud?
Configuration parameters (including the API Key) are stored only locally in ~/.codebuddy/models.json and are not uploaded to the cloud. Models manually configured through that file in the early days can still be used normally after the interface upgrade, and can be viewed, edited, and deleted in the interface—no need to touch configuration files for daily management. When you stop using a model, the official recommendation is to clean its Key out of models.json.
You should see: after saving, a custom model group appears in the model selector of the conversation interface (showing model names and provider icons); select one to start using it.
- I added a custom model and saw it in the selector's custom group
Step 5: Connect a local model via Ollama (no API Key needed)
Why: You don't want sensitive content leaving your machine, your intranet environment has no external internet access, or you just want a zero-cost trial—local Ollama solves all three at once.
What is Ollama (official wording): an open-source tool for running large models locally. After installation, you can pull and run open-source models with a single command; once started, it listens on a local HTTP port (default 11434) and automatically provides an OpenAI-protocol-compatible API for WorkBuddy to connect to.
Three types of applicable scenarios listed officially:
| Scenario | Benefit |
|---|---|
| Data privacy / intranet compliance | Code and conversation content never leave your machine—suitable for finance, government/enterprise, and sensitive projects |
| Zero-cost trial | No Token consumption, no paid API Key—use your hardware as the compute power |
| Works offline | Usable on airplanes, intranet dev machines, and environments without internet, so you can still use WorkBuddy |
Steps:
- Install and start Ollama, and pull an open-source model (follow the official Ollama instructions for installation and model pulling), keeping the local service running
- WorkBuddy Settings → Models → Add, select Ollama from the provider list—the API address is filled in automatically (Ollama listens on local port 11434 by default)
- Enter the exact name of the installed model and save
- No API Key needed—the official statement says "models pointing to the local machine (such as local Ollama) do not require an API Key"
⚠️ A local model ≠ the entire task stays offline
A local model only guarantees that "model inference" happens on your machine. The official international documentation additionally warns: web search, external connectors, or other online services may still access the network—connecting a local model does not mean all data for the task will stay on your computer, nor does it guarantee compliance with specific regulatory requirements. For compliance-sensitive scenarios, first think clearly about whether the task will touch the network; for an in-depth discussion of security mechanisms, see 5.4 Data Security and Privacy.
You should see: the local model you configured appears in the custom group of the model selector; after you send a message, it is answered by the local model, and local resource usage rises noticeably (your hardware is your compute power).
- My local model appeared in the selector, and I was never asked for an API Key
Step 6: Troubleshooting 4100: What to do if a custom model is grayed out / requests aren't sent
Why: The most common failure for custom models is just one—incomplete configuration. Learn to recognize it, and it can be fixed in 10 seconds.
Symptoms (official error code 4100):
- In the model selector: the model is grayed out and cannot be selected, with an error icon (e.g., the model name was written by a third-party tool but not fully configured)
- Hover tooltip: "The current custom model has no API address or API key configured. Please go to settings and configure it before use"
- If a request is still sent: the conversation shows the same message and provides a "Go to configuration" entry, error code 4100, and the request is not sent
The fix (3 steps):
- Click "Go to configuration" in the prompt (available in both the hover card and the conversation message) to open that model's configuration page directly
- Fill in the API address and API Key and save
- Done—official wording: "after completing and saving, the model can be selected normally" and "after completing and saving, the model becomes available again without a restart"
Exception: models pointing to the local machine (such as local Ollama) do not require an API Key and are not subject to this restriction.
For other error codes (3002 network connection, 3008 no model response, 6003, etc.), see the official "Error Code Handling Guide" and this site's Troubleshooting Quick Reference: Error Codes and Frequent Issues.
You should see: after completing and saving, the model changes from grayed out to selectable; send another message and the request goes out normally.
- I know that 4100 = the custom model is missing its API address or API Key, and no restart is needed after completing them
For detailed steps, refer to the official documentation: Model Configuration, Error Code Handling Guide.
Checkpoints ✅
- Can switch between the Fast / Balanced / Ultimate tiers in the model selection panel and describe the task types each is suited for
- Can describe the purpose of the "Vision" tag and find the model's credit multiplier notation (x0.06~x1.62)
- Knows that a larger context window means higher per-turn consumption, that tiers only apply to the current task, and that Economy mode is recommended to stay on
- Has added a custom model (any of Token Plan / Coding Plan / custom API / Ollama) and knows the configuration is stored only in the local models.json and never uploaded to the cloud
- When encountering 4100, clicks "Go to configuration" to complete the setup instead of suspecting the software is broken
Pitfall Reminders
❌ Maxing out the context window, assuming bigger is always better
Symptom: The context tier is set to maximum, task quality doesn't change, and credits drain rapidly.
Cause: Every turn of the conversation sends the full current context to the model; the larger the window, the higher the per-turn consumption. The official statement: "A larger window doesn't make the model smarter—it only multiplies the consumption per turn."
Correct approach: Keep the default tier; only increase it when handling extra-long documents or when the model needs to fully remember a large amount of history—and the tier only applies to the current task, so there's no need to deliberately switch back; new tasks automatically revert to the default.
❌ Treating an API Key as ordinary text—sharing it or committing it to a public repository
Symptom: Posting a screenshot of the configuration with the Key to a group chat, or uploading models.json to a code repository, and the quota gets used up by someone else.
Cause: The API Key is tied to your account and plan quota. Official security notice: leakage may allow others to misuse your quota, incur extra charges, or create data security risks.
Correct approach: Enter the Key only in WorkBuddy's local configuration (models.json is never uploaded to the cloud); never fill it in or export it in shared environments; when it's no longer used, remove it from models.json.
❌ Sending messages anyway with a grayed-out custom model, assuming the software is broken
Symptom: Selecting a grayed-out custom model and sending a message, the conversation reports 4100, and the first instinct is to reinstall.
Cause: The model has no API address or API Key configured—WorkBuddy simply doesn't send the request; this is a protection mechanism, not a malfunction.
Correct approach: Click "Go to configuration" in the prompt, fill in the API address and API Key, and save—no restart needed. Local Ollama models don't require a Key and are unaffected.
❌ Assuming "connected a local model = fully offline, data absolutely never leaves the machine"
Symptom: After connecting Ollama for compliance, you let WorkBuddy handle everything without worry.
Cause: A local model only guarantees that model inference happens on your machine; online features such as web search and external connectors still access the network (per the official international documentation).
Correct approach: For compliance-sensitive scenarios, first confirm whether the task will touch the network (e.g., don't ask it to search the web within a local-model task); for security boundaries, see 5.4 Data Security and Privacy.
Advanced Tips (Optional)
🔧 Custom protocol: only touch this when the API goes through a gateway / proxy. When the model service uses a non-standard URL path (wrapped by a gateway or proxy layer), enable the "Custom protocol" toggle in advanced configuration: Off (default) = use the standard /chat/completions path, with automatic validation and completion of the API address; On = send requests directly to the URL you entered, skipping path validation and auto-completion. Generally keep it off unless the service provider explicitly requires it.
🔧 How to view models on mobile. The mobile app can show the model list, capability tags, and multipliers, and can also adjust thinking intensity (e.g., Low / Medium / High / Ultra) and the context window (e.g., 300K / 1M)—the entry is the configuration entry at the top-right of the model selection panel. However, the mobile side does not provide an entry for "configuring custom models", and the model list only shows available models; on the mini-program side, model selection only appears after the first message of a task is sent.
🔧 Using the international version of workbuddy.ai? The configuration entry point and file paths differ: Avatar → System Settings → Models, the configuration file is at ~/.workbuddy-ai/models.json, and the default Ollama endpoint address officially listed is http://localhost:11434/v1/chat/completions. For international version differences, see 5.5 International Version of workbuddy.ai: Differences from the Domestic Version and How to Connect.
Lesson Summary
✅ Two-tier models: WorkBuddy models come in 14 built-in options (Hunyuan / GLM / Kimi / MiniMax / DeepSeek) and custom models — built-in ones work out of the box and consume credits at a rate multiplier; custom ones you connect yourself, don't consume WorkBuddy credits, and fees are paid to third parties.
✅ Three modes: Fast (simple tasks, fastest and cheapest) / Balanced (the vast majority of tasks — choose this if unsure) / Ultimate (complex tasks, more stable but more expensive).
✅ Choosing a built-in model: Look at capability tags (Reasoning / Vision) + credit multiplier (x0.06~x1.62, lower is cheaper) + load status (green / orange / red / unavailable); if unsure, use the default Auto and the system will match automatically.
✅ Two parameters: Keep the thinking intensity at its default; the larger the context window, the more expensive each turn, and the tier only applies to the current task; keep Economy Mode on (Settings → Agent) to save credits.
✅ Custom integration: For provider presets (Tencent Cloud Token Plan / Coding Plan / Custom API), just fill in the API Key; local Ollama needs no Key (default 11434 port); if it's not in the presets, choose Custom and fill in manually. Configuration is stored only locally in ~/.codebuddy/models.json and never uploaded to the cloud.
✅ 4100: Custom model missing endpoint URL or API Key → click "Go to Configuration" to fill in and save; it recovers immediately with no restart needed.
Key configuration recap:
| Configuration Item | Recommended Value | Notes |
|---|---|---|
| Model Mode | Balanced | For the vast majority of tasks; switch to Fast for simple tasks, or Ultimate for complex ones |
| Built-in Model | Auto, or pick by multiplier + tag | Use a low-multiplier reasoning model for simple tasks; choose the "Vision" tag for image understanding |
| Thinking Intensity | Default | The official default is already suited to the vast majority of tasks |
| Context Window | Default tier | The larger it is, the more expensive each turn; increase it only for extra-long documents, and the tier applies only to the current task |
| Economy Mode | On | Settings → Agent → Economy Mode; automatically compresses context and saves credits in long conversations |
| Custom Model | Prefer provider presets | URL and capability tags are auto-filled; just fill in the API Key |
| Local Ollama | Choose Ollama as the provider | Address is auto-filled (listens on the default 11434 port); no API Key needed |
Frequently Asked Questions
WorkBuddy How do I connect a third-party API?
Open the "Models" section on the settings page. When adding a model, choose "Custom API" from the provider list — it has built-in preset entries for mainstream vendors, enabling one-click connection. The URL, model list, and capability tags are filled in automatically; you only need to supply the API Key to save and use it. If the service is not in any preset list, select "Custom", and manually fill in the URL, API Key, and model name. The configuration is stored only locally in ~/.codebuddy/models.json and is not uploaded to the cloud.
WorkBuddy What are the steps to connect an API?
There are four steps in total: open the "Models" section on the settings page; add a model and select from the provider list (Tencent Cloud Token Plan, Coding Plan, Custom API, or Ollama, etc.); complete the information according to the chosen method (preset methods usually only require the API Key; for Token Plan, you also need to first select the plan type and then a supported model); after saving, the configuration is automatically persisted, and the model appears in the custom group of the model selector. The whole process requires no manual editing of configuration files.
WorkBuddy How do I configure a custom model?
Add it in Settings → Models: when you select a provider preset, the API endpoint and capability tags are filled in automatically, and you only need to supply the API Key; local Ollama does not require an API Key — after selecting Ollama as the provider, the API endpoint is filled in automatically (listening on the local 11434 port by default), then enter the exact name of an installed model and save. Note that the mobile app does not offer an entry point for "configuring custom models"; configuration must be done on the desktop client. When you no longer use a model, the official recommendation is to clean up its key in models.json.
WorkBuddy How do I choose a model?
First select one of the three tiers in the model selection panel: Fast for simple tasks, Balanced for the vast majority of tasks, and Ultimate for complex tasks — the official recommendation is to use Balanced when unsure which tier to pick. Then choose a specific built-in model based on capability tags, credit multipliers (x0.06~x1.62, lower is cheaper), and busy status; if unsure which one to pick, use the default Auto (smart automatic selection), and the system will automatically match the optimal model.
WorkBuddy Are API interfaces supported?
Yes. 14 built-in models work out of the box; for models outside the built-in list, connect a third-party API via the "Models" section on the settings page (API Key, Tencent Cloud Token Plan, Coding Plan), or use local Ollama to connect open-source models. Custom models do not consume WorkBuddy credits; any costs incurred are paid by you directly to the third party.
Next Lesson Preview
Now that you know the models and multipliers, the next lesson is about the money. In the next lesson you'll learn about 2.6 Credits and Pricing: Free Allowance, Four Subscription Tiers, and Money-Saving Tips.
You will learn:
- How many base credits the trial edition gives per month (500), and how much each of the four paid tiers provides
- Credits are issued monthly, don't roll over after expiry, and deductions use the earliest-expiring credits first
- How to buy add-on packs and how to request refunds
- What money-saving tips exist from official guidance and third-party testing