Model configuration

Xiajiao supports any OpenAI-compatible API. This page covers major providers. If you have not run the app yet, start with Quick start. After models work, tune personas and tokens with the SOUL.md guide.

Global settings—theme, language, and default LLM model

Agent card—dropdown to assign a model per agent

Where to configure

After login:

Settings → Model management → Add configuration

Each provider needs:

Field	Notes
Name	Any label (e.g. “Qwen”)
API base URL	Provider endpoint
API key	Secret from vendor
API type	openai-completions or anthropic-messages
Default model	Default model id for that provider

OpenAI

Field	Value
API base URL	https://api.openai.com/v1
API type	openai-completions
Keys	platform.openai.com/api-keys

Suggested models

Model	Notes	Price
gpt-4o	Flagship, multimodal	$5/M in, $15/M out
gpt-4o-mini	Lightweight, good value	$0.15/M in, $0.6/M out
gpt-4-turbo	Previous flagship	$10/M in, $30/M out
o1	Reasoning-focused	$15/M in, $60/M out

Practices

• Daily chat: gpt-4o-mini
• Hard tasks: gpt-4o
• Code: gpt-4o

Anthropic (Claude)

Field	Value
API base URL	https://api.anthropic.com
API type	anthropic-messages
Keys	console.anthropic.com

API type

Claude must use anthropic-messages, not openai-completions.

Suggested models

Model	Notes	Price
claude-sonnet-4-20250514	Strong code and reasoning	$3/M in, $15/M out
claude-3-5-haiku-20241022	Fast, smaller	$1/M in, $5/M out
claude-opus-4-20250514	Maximum reasoning	$15/M in, $75/M out

Practices

• Code: Claude Sonnet
• Long writing: Sonnet (large context)
• Budget: Haiku

Qwen (Alibaba DashScope)

Field	Value
API base URL	https://dashscope.aliyuncs.com/compatible-mode/v1
API type	openai-completions
Keys	dashscope.console.aliyun.com

Suggested models

Model	Notes	Price
qwen-max	Flagship	CN¥20/M in, CN¥60/M out
qwen-plus	Balanced	CN¥0.8/M in, CN¥2/M out
qwen-turbo	Fast, cheap	CN¥0.3/M in, CN¥0.6/M out
qwen-long	Long context	CN¥0.5/M in, CN¥2/M out

Practices

• Daily: qwen-turbo
• Harder work: qwen-plus
• Best quality: qwen-max

New accounts

Qwen often offers free credits for new signups—check the vendor site.

DeepSeek

Field	Value
API base URL	https://api.deepseek.com
API type	openai-completions
Keys	platform.deepseek.com

Suggested models

Model	Notes	Price
deepseek-chat	General chat	CN¥1/M in, CN¥2/M out
deepseek-coder	Code	CN¥1/M in, CN¥2/M out
deepseek-reasoner	Reasoning	CN¥4/M in, CN¥16/M out

Practices

• Strong price/performance vs flagship Western models
• Code: deepseek-coder
• Chat: deepseek-chat

Kimi (Moonshot)

Field	Value
API base URL	https://api.moonshot.cn/v1
API type	openai-completions
Keys	platform.moonshot.cn

Suggested models

Model	Notes	Price
moonshot-v1-8k	8K context	CN¥12 / M tokens
moonshot-v1-32k	32K	CN¥24 / M tokens
moonshot-v1-128k	128K	CN¥60 / M tokens

Practices

• Default: 8K
• Long docs: 128K

GLM (Zhipu)

Field	Value
API base URL	https://open.bigmodel.cn/api/paas/v4
API type	openai-completions
Keys	open.bigmodel.cn

Suggested models

Model	Notes	Price
glm-4-plus	Flagship	CN¥50 / M tokens
glm-4-flash	Fast	Free
glm-4-long	Long text	CN¥1 / M tokens

Free tier

glm-4-flash is free—good for experiments and light use.

Ollama (local)

Field	Value
API base URL	http://localhost:11434/v1
API type	openai-completions
API key	Omit or use a placeholder like ollama

Install Ollama

# macOS / Linux
curl -fsSL https://ollama.com/install.sh | sh

# Windows: installer from https://ollama.com/download

Pull models

ollama pull llama3.1          # Llama 3.1 8B
ollama pull qwen2.5           # Qwen 2.5
ollama pull mistral           # Mistral 7B
ollama pull codellama         # Code-focused
ollama pull deepseek-coder-v2

Hardware

Size	Min VRAM	Comfortable VRAM
7B	4GB	8GB
13B	8GB	16GB
70B	40GB	48GB+

CPU-only works but is slow.

Practices

• Free, private, offline-friendly
• Privacy-sensitive workloads
• ~8B models run well on consumer GPUs
• Chinese: qwen2.5 is a solid default

OpenRouter

Field	Value
API base URL	https://openrouter.ai/api/v1
API type	openai-completions
Keys	openrouter.ai/keys

One key routes to many models—handy if you switch often.

Example model ids

openai/gpt-4o
anthropic/claude-3.5-sonnet
google/gemini-pro-1.5
meta-llama/llama-3.1-70b-instruct

Multiple providers

Xiajiao can keep several providers and assign different models per agent.

Example mix

Agent	Provider	Model	Rationale
🤖 Xiajiao steward	Qwen	qwen-turbo	Simple ops, low cost
✍️ Novelist	Claude	claude-sonnet	Quality writing
📝 Editor	DeepSeek	deepseek-chat	Cheap text work
🌐 Translator	OpenAI	gpt-4o	Strong multilingual
💻 Coding assistant	Claude	claude-sonnet	Strong code

Budget-first

Agent	Provider	Model	Rough monthly
All	Qwen	qwen-turbo	< CN¥5

Quality-first

Role	Provider	Model	Rough monthly
Creative	Claude	claude-sonnet	~$10
Tools	OpenAI	gpt-4o	~$10

Free stack

Agent	Provider	Model
All	Ollama	qwen2.5 / llama3.1

Cost tips

Principles

1. Match model to task—not everything needs GPT-4o or Claude Opus
2. Cheap models for translation, summary, format tweaks; premium for creation, code, hard reasoning
3. Tighter SOUL.md → fewer prompt tokens each call

Example cost for one “500-word tech post” prompt

Model	Input tok	Output tok	Rough cost
GPT-4o	~800	~600	~$0.012
Claude Sonnet	~800	~600	~$0.009
DeepSeek Chat	~800	~600	~CN¥0.004
Qwen Turbo	~800	~600	~CN¥0.003
Ollama	~800	~600	CN¥0

Example routing

Steward (ops) → Qwen Turbo
Translator → DeepSeek Chat
Coding assistant → Claude Sonnet
Casual agent → Ollama qwen2.5

Reduce token usage

1. Shorten SOUL.md—roughly ~1 token per word saved
2. Limit memory injection: e.g. AUTO_MEMORY_TOP_K=3
3. Disable unused tools—each tool adds ~100–200 tokens in definitions
4. Start fresh threads when history grows too long

Troubleshooting

Invalid API key

Symptom: 401 Unauthorized

Fix: Re-copy the key (no spaces), confirm it is active.

Wrong base URL

Symptom: ECONNREFUSED or 404

Fix: Check trailing paths:

• Wrong: https://api.openai.com (missing /v1)
• Right: https://api.openai.com/v1
• Wrong: http://localhost:11434 (Ollama needs /v1)
• Right: http://localhost:11434/v1

Wrong model name

Symptom: model not found

Fix: Match vendor spelling exactly—case-sensitive.

Claude fails

Symptom: errors from Anthropic

Fix: Type must be anthropic-messages, not openai-completions.

Ollama connection refused

Symptom: ECONNREFUSED

Fix:

1. ollama list — daemon running
2. Default port 11434
3. Remote Ollama: bind 0.0.0.0 if Xiajiao runs on another host

Quick verification checklist

✅ OpenAI-compatible URLs end with /v1 where required
✅ API key has no spaces or line breaks
✅ Model id matches provider docs
✅ API type correct (Anthropic → anthropic-messages, else openai-completions)
✅ Test agent sends one message successfully

Related docs

• Quick start — install to first reply
• SOUL.md guide — personas and token use
• Multi-agent chat
• Performance
• Glossary
• FAQ