# Models and providers: all 16 providers compared

> NearWork reaches 16 model providers over the OpenAI, Anthropic and Gemini protocols. Full provider table with endpoints, how the seven model slots work, and what changes when you bring your own API key.

NearWork does not train models; it is the layer that connects them to work on your machine. Which provider you use, which model fills which slot, and whether requests go through official credits or your own key are all settings. There are **16 third-party providers** built in, all reached over the OpenAI, Anthropic or Gemini protocol — if a service speaks one of those, it can be added manually even when it is not on the list.

## Provider table

| Provider | Region | Protocol | Default endpoint |
|---|---|---|---|
| DeepSeek | China | OpenAI | `https://api.deepseek.com` |
| Qwen (Alibaba Cloud) | China | OpenAI | `https://dashscope.aliyuncs.com/compatible-mode/v1` |
| Kimi (Moonshot) | China | OpenAI | `https://api.moonshot.cn/v1` |
| Zhipu GLM | China | OpenAI | `https://open.bigmodel.cn/api/paas/v4` |
| Doubao (Volcengine Ark) | China | OpenAI | `https://ark.cn-beijing.volces.com/api/v3` |
| MiniMax | China | OpenAI | `https://api.minimaxi.com/v1` |
| ERNIE (Baidu Qianfan) | China | OpenAI | `https://qianfan.baidubce.com/v2` |
| Tencent Hunyuan | China | OpenAI | `https://tokenhub.tencentmaas.com/v1` |
| StepFun | China | OpenAI | `https://api.stepfun.com/v1` |
| Xiaomi MiMo | China | OpenAI | `https://api.xiaomimimo.com/v1` |
| OpenAI | United States | OpenAI | `https://api.openai.com/v1` |
| Anthropic | United States | Anthropic | `https://api.anthropic.com` |
| Gemini | United States | Gemini | `https://generativelanguage.googleapis.com` |
| xAI Grok | United States | OpenAI | `https://api.x.ai/v1` |
| Groq | United States | OpenAI | `https://api.groq.com/openai/v1` |
| Mistral | Europe | OpenAI | `https://api.mistral.ai/v1` |

The endpoints above are defaults, not facts about your setup: the request always goes to whatever `baseURL` you configure. Self-hosted gateways, corporate relays and third-party proxies all work — keep the protocol, change the address and the model ID.

## Seven slots instead of one model for everything

| Slot | Used for | Typical work |
|---|---|---|
| Fast | Everyday Q&A and rewriting | The bulk of your turns: quick and cheap |
| Think | Hard reasoning | Planning, stubborn bugs, long chains of analysis |
| Vision | Reading images | Screenshots, charts, scanned pages |
| Image edit | Generating and editing images | Posters, diagrams, restyling |
| Speech generation | Text to speech | Reading replies aloud, voice-over |
| Speech recognition | Speech to text | Voice input, meeting transcripts |
| Embedding | Semantic retrieval | Memory and knowledge-base recall |

Splitting the choice keeps cost and quality where they belong: a cheap fast model for everyday turns, the expensive one reserved for thinking, vision and speech billed only when images or audio are actually involved. In chat you only switch Fast / Think above the input box; the other slots are consumed by the matching feature.

An empty slot means **that capability is unavailable**, never a silent substitution — attach an image without a vision model and you get an explicit "not configured" message.

## Official credits versus your own API key

| Route | Path of a request | Who bills you | Suits |
|---|---|---|---|
| NearWork credits | Through the official relay, authenticated and metered | Your NearWork plan (credits) | Anyone who wants zero setup |
| Your own API key | Straight to the provider's endpoint | The provider | People with accounts, or strict data-path requirements |

You can switch at any time, and different sessions may use different routes. Note that **a personal key does not replace signing in** — a NearWork account is still required. On desktop, keys are encrypted with the OS facility (Electron `safeStorage`) before they hit disk, so expect to re-enter them after moving machines.

## Three practical selection rules

1. **Match the network**: DeepSeek, Qwen, Kimi, Zhipu and Doubao have low latency from mainland China and are enough for daily tasks.
2. **Judge coding and long context by trial, not leaderboards**: run two or three of your own real tasks; the gap between models on your prompts is often bigger than the ranking gap.
3. **For sensitive content, use direct connections or local models**: the official relay sees request bodies; with your own key the path is provider-direct.

## Frequently asked questions

### Is the model list hard-coded?

The provider list is built in, but model IDs are not restricted: you type whatever ID the provider publishes into each slot, and new releases usually need no client update.

### Can I connect a provider that is not listed?

Yes. Add a provider entry, pick a compatible protocol (OpenAI-compatible covers most of the market), then fill in `baseURL`, `apiKey` and the model IDs.

### Does switching models lose context?

No. Changing slots or providers inside a session keeps the history; only subsequent replies come from a different model.

## Next steps

- First time here? Start with the [quick start](/en/docs/start/quick-start).
- Want to know what those models can actually call? See the [tool inventory](/en/docs/features/tools).
- Care about keys and data paths? Read [Data and privacy](/en/docs/faq/data-privacy).
