Multi-provider routing

Every /ai/chat request passes through detectProvider(model) which selects the backend:

Model pattern	Provider	Auth required
meta/*, mistral/*, qwen/*, deepseek/*, moonshot/*, etc.	Workers AI	None (edge-bound)
meta-llm	Workers AI (Llama 3.1 8B alias)	None
gpt-*, o3, o4-mini	OpenAI	OPENAI_API_KEY secret or BYOK
claude-*	Anthropic	ANTHROPIC_API_KEY secret or BYOK

Auto-routing (auto)

When model: "auto", Neureus classifies the task type and selects the best available model:

Task type	Preferred model
coding	qwen/coder-32b → gpt-4o
reasoning	deepseek/r1-32b → claude-opus-4-8
summarization	claude-haiku-4-5 → meta/llama-3.1-8b
extraction	gpt-4o-mini → meta/llama-3.1-8b
chat	meta/llama-3.3-70b → gpt-4o-mini

Selection falls through to the next provider if the preferred one has no key configured. Workers AI models always work — no key required.

Available model IDs

Workers AI (no key required):

• meta/llama-3.3-70b — general chat, fast
• meta/llama-3.1-8b — smallest, cheapest
• meta/llama-3.2-3b — ultra-fast edge inference
• meta/llama-4-scout-17b — Llama 4, long context
• qwen/coder-32b — code generation
• qwen/qwq-32b — reasoning
• deepseek/r1-32b — reasoning tasks
• mistral/7b — multilingual
• mistral/small-24b — balanced multilingual
• moonshot/kimi-k2 — long-context general
• moonshot/kimi-k2-code — long-context coding
• google/gemma-4-27b — general purpose

OpenAI (BYOK or global key):

• gpt-4o, gpt-4o-mini, o3, o4-mini

Anthropic (BYOK or global key):

• claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5

meta-llm vs auto

Alias	Behaviour
meta-llm	Hardcoded to meta/llama-3.1-8b — no classification, always the same model
auto	Runs resolveAutoModel() — classifies the prompt, selects across all providers

Use auto for production smart-routing. meta-llm exists for backwards compatibility and quick testing.

Streaming

All providers return normalized OpenAI-compatible SSE. The streamLlmComplete function translates provider-specific formats transparently.