Model Providers en

🌐 简体中文 · English · 📖 README · 🏠 Wiki home

🤖 Model Providers

PandaGPT ships with 15 model providers, covering the major international and Chinese platforms, plus any OpenAI-compatible endpoint. Switching providers is a one-line change and a reload.

Overview

provider id	Provider	Wire format	Default model	Image input by default
openai	OpenAI	OpenAI	gpt-4.1-mini	✅
gemini	Google Gemini	Native Gemini	gemini-3.8-flash	✅
anthropic	Anthropic Claude	Native Claude Messages	claude-sonnet-5-5	✅
azure	Azure OpenAI	OpenAI (api-key header)	deployment name	✅
xai	xAI Grok	OpenAI-compatible	grok-4	✅
mistral	Mistral AI	OpenAI-compatible	mistral-medium-latest	✅
groq	Groq	OpenAI-compatible	openai/gpt-oss-120b	❌
openrouter	OpenRouter (hundreds of models)	OpenAI-compatible	openai/gpt-4.1-mini	✅
deepseek	DeepSeek	OpenAI-compatible	deepseek-chat	❌
qwen	Alibaba Cloud Model Studio · Qwen	OpenAI-compatible	qwen-plus	❌
moonshot	Moonshot AI · Kimi	OpenAI-compatible	kimi-k2-turbo-preview	❌
zhipu	Zhipu AI · GLM	OpenAI-compatible	glm-4.6	❌
doubao	Volcengine Ark · Doubao	OpenAI-compatible	doubao-seed-1-6-250615	✅
siliconflow	SiliconFlow	OpenAI-compatible	deepseek-ai/DeepSeek-V3	❌
custom	Any OpenAI-compatible endpoint	OpenAI-compatible	required	❌

⚠️ Model names change as providers update; the table lists defaults only. Use a model your account can actually access, as shown in your provider's console.

Configuration

Each provider has its own block under panda.model_selection, so you can keep keys for several providers and switch at any time:

panda:
  model_selection:
    provider: 'anthropic'          # the provider in use
    anthropic:
      api_key: 'sk-ant-...'
      api_url: 'https://api.anthropic.com/v1/messages'
      model: 'claude-sonnet-5-5'
      vision: true                 # does the model accept images?
    deepseek:
      api_key: 'sk-...'
      api_url: 'https://api.deepseek.com/chat/completions'
      model: 'deepseek-chat'
      vision: false

Field	Meaning
api_key	The key. PANDAGPT_<ID_IN_UPPER_CASE>_API_KEY also works and wins when non-empty
api_url	Endpoint. Must be HTTPS with no query, fragment or URL credentials. Empty means the default
model	Model name: letters, digits and ._:/- only
vision	Whether vision captures are sent as images. When off, the agent receives text-only observations (object IDs, blocks, distances, entities). Vision still works, just without pixels

Provider notes

OpenAI

Uses Chat Completions with max_completion_tokens. Parallel tool calls and image input are supported.

Google Gemini

%s in the URL is replaced with the model name. The key is sent in the x-goog-api-key header. Gemini 3 requires thought signatures (thoughtSignature) to be echoed back on multi-turn function calls, and the plugin stores and returns them automatically.

Anthropic Claude

Uses the native Messages API (x-api-key + anthropic-version headers), not a compatibility layer. Consecutive tool results are merged into one user message to satisfy Claude's strict role alternation, and thinking blocks are never shown to players.

Azure OpenAI

api_url is required, for example https://<resource>.openai.azure.com/openai/v1/chat/completions. Set model to your deployment name. The key is sent in the api-key header.

Chinese providers (DeepSeek, Qwen, Kimi, GLM, Doubao, SiliconFlow)

All use the OpenAI-compatible API; the plugin automatically sends max_tokens instead of max_completion_tokens.

• Doubao: model may be a model name or a Volcengine Ark endpoint ID (ep-...).
• Qwen: for images, switch to a vision model (e.g. qwen-vl-plus) and set vision: true.
• DeepSeek: deepseek-chat supports tool calling; check the official docs for reasoning models.

OpenRouter, Groq, xAI, Mistral

OpenRouter gives one key access to hundreds of models; write model names as vendor/model, e.g. anthropic/claude-sonnet-5-5. Groq is known for very fast inference, good for latency-sensitive servers.

Custom compatible endpoint custom

For any service implementing OpenAI Chat Completions with tools, such as a self-hosted vLLM, LiteLLM, One API or New API gateway. Both api_url and model are required, and the URL must be HTTPS.

Choosing a model

Need	Suggestion
Full experience (including images)	A model with vision: true and reliable tool calling, e.g. Claude, GPT, Gemini, Doubao
Lowest cost	DeepSeek, Qwen, GLM and similar; with vision off, vision still works as text
Mainland China networks	Chinese providers connect directly; international ones can use panda.proxy (HTTP CONNECT)
Speed	Groq, or any provider's flash / mini / turbo tier

💡 One turn may call the model several times (up to 8 steps by default). Control cost with panda.agent.max_steps and max_tool_calls; see Configuration.

Networking and retries

• Each request has an overall timeout (panda.requests.timeout_seconds) covering retries and reading the body.
• Only HTTP 429/502/503/504 and connection failures are retried, honouring Retry-After. A retry only re-fetches the model reply and never re-runs tools.
• Response bodies are capped (1 MiB by default); errors never include raw responses, URLs or keys.
• Proxy: panda.proxy supports HTTP CONNECT proxies. With a username and password, add -Djdk.http.auth.tunneling.disabledSchemes= to the JVM flags; the plugin warns in the console if it is missing.

---

⬅️ 📦 Installation  ·  📚 Contents  ·  💬 Talking to Panda ➡️