Model Router

IO runs a smart multi-model router. You specify "model": "auto" and IO picks the best model for your query. Or lock to a specific model — your choice.

The roster spans both open-weight models (self-hosted, zero data-sharing) and frontier models (GPT, Gemini, Claude) — all routed through the same zero-retention inference layer. Every response carries a signed receipt regardless of which model answered.

Text & Code Models

ID	Model	Provider	Context	Best For
`dolphin`	Dolphin-Mistral-24B	Venice (open-weight)	32K	Flagship — uncensored general purpose
`deepseek`	DeepSeek	Open-weight	128K	Code, math, reasoning
`qwen`	Qwen	Open-weight	32K	Multilingual, translation
`llama`	Llama 3.1 70B	Open-weight	128K	Long-context, research
`mistral`	Mistral	Open-weight	32K	Fallback, fast
`gpt`	GPT-4o	OpenAI	128K	Frontier reasoning, vision
`gemini`	Gemini 2.0 Flash	Google	1M	Ultra-long context, multimodal
`claude`	Claude Sonnet	Anthropic	200K	Deep reasoning, safety-aware
`auto`	Router decides	—	—	Smart routing in <3s

Image Generation Models

ID	Model	Provider	Best For
`imagen`	Imagen 3	Google (Gemini)	Photorealism, fine detail
`dalle`	DALL·E 3	OpenAI (GPT)	Creative, prompt-faithful
`sdxl`	Stable Diffusion XL	Open-weight	Fast, highly customisable

Auto Router

When you send "model": "auto", the router analyzes your prompt and selects the best model:

Code / math / reasoning → DeepSeek or GPT-4o
Non-English → Qwen
Ultra-long context (>128K tokens) → Gemini 2.0 Flash (1M window)
Complex analysis / nuanced reasoning → Claude Sonnet
General uncensored → Dolphin-Mistral-24B
Fallback / low latency → Mistral

The router decides in under 3 seconds and names the chosen model in the response receipt.

Auto-router decision tree — how IO picks the right model for your prompt.

Specifying a Model

{"model": "deepseek", "messages": [{"role": "user", "content": "Write a Rust sorting algorithm"}]}

{"model": "claude", "messages": [{"role": "user", "content": "Analyze the risks in this contract"}]}

{"model": "gemini", "messages": [{"role": "user", "content": "Summarize this 500-page document"}]}

Use the model ID from the registry above.

Broadcast Mode

Send to multiple models simultaneously:

{"models": ["dolphin", "gpt", "claude"], "messages": [{"role": "user", "content": "What do you disagree on?"}]}

Diff Mode

Compare two models side-by-side:

{"model_a": "gpt", "model_b": "claude", "messages": [{"role": "user", "content": "Explain consciousness"}]}

Returns a divergence score + structured key differences.

Diffuse Mode

Broadcast to multiple models, then synthesize unique insights into one answer:

{"model": "diffuse", "models": ["dolphin", "gpt", "gemini", "claude"], "messages": [{"role": "user", "content": "What is the future of AI?"}]}

The response merges non-overlapping insights with per-model attribution.

Pricing

Each model has a per-token rate. auto routing applies the selected model's rate. Check current pricing at /v1/models.

All billing is on-chain via $IO token — no credit cards, no Stripe, no privacy leaks.