Expanding Jibun Corp's AI Hub to 20 Providers — Including Cerebras Ultra-Fast Inference & Chinese AI

#flutter #supabase #buildinpublic #ai

Expanding Jibun Corp's AI Hub to 20 Providers

What Changed

The ai-hub Supabase Edge Function now supports 20 AI providers, up from 14. The six new additions:

Provider	Highlight	Default Model
Cerebras	Ultra-fast (1000+ tokens/s)	llama-3.3-70b
NVIDIA NIM	Enterprise inference API	meta/llama-3.1-70b-instruct
Moonshot (Kimi)	Long context, Chinese-origin	moonshot-v1-8k
AI21 Labs	Jamba SSM+Transformer hybrid	jamba-1.5-mini
01.AI (Yi)	Open-source Chinese LLM	yi-lightning
Zhipu AI (GLM)	GLM-4 family, generous free tier	glm-4-flash

The OpenAI-Compatibility Pattern

Every provider except Google Gemini uses the same OpenAI-compatible body builder:

const OPENAI_COMPAT_BODY = (messages: unknown[], model: string) => ({
  model,
  messages,
  max_tokens: 1000,
  temperature: 0.7,
});

cerebras: {
  envKey: "CEREBRAS_API_KEY",
  chatUrl: "https://api.cerebras.ai/v1/chat/completions",
  defaultModel: "llama-3.3-70b",
  buildBody: OPENAI_COMPAT_BODY,
  parseResponse: OPENAI_COMPAT_PARSE,
},

This pattern means adding a new OpenAI-compatible provider is literally ~7 lines of config.

Why Cerebras Stands Out

Cerebras uses purpose-built silicon (the Wafer Scale Engine) for inference. The result: the same Llama-3.3-70b model runs 10x+ faster than on GPU clusters. For real-time chat apps, this matters.

Provider Status in the Flutter UI

Each provider maps to a status in AiProviderStatus:

implemented — API key set, ready to use
apiKeyRequired — Code is ready, waiting for the Supabase Secret
paidPlanRequired — Needs a paid plan subscription
notImplemented — Not yet coded

All 6 new providers are now apiKeyRequired — code complete, secrets pending.

What's Next

The BYOK (Bring Your Own Key) flow lets users supply their own API keys via the AI Provider Status page. As more keys get added, the green implemented count grows.

Building in public: https://my-web-app-b67f4.web.app/