API reference

Both endpoints are OpenAI-compatible, so any OpenAI SDK, LiteLLM, LangChain or plain HTTP client works. The endpoint is the same for everyone — only your key differs.

Chat completions — dhandare-4-mini

POST https://models.quantforge.co.in/v1/chat/completions

curl https://models.quantforge.co.in/v1/chat/completions \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "dhandare-4-mini",
    "messages": [
      {"role": "system", "content": "You are a health-check probe. Reply with exactly the word: OK"},
      {"role": "user", "content": "Reply with exactly the word: OK"}
    ]
  }'

The x-api-key header is accepted too, for Anthropic-style clients.

Embeddings — dhandare-embed-1

POST https://embeddings.quantforge.co.in/v1/embeddings

curl https://embeddings.quantforge.co.in/v1/embeddings \
  -H "Authorization: Bearer YOUR_EMBED_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "dhandare-embed-1", "input": "text to embed"}'

Returns 384-dimensional vectors, L2-normalised. input accepts a string or an array of strings.

Usage and billing in the response

"usage": {
  "prompt_tokens": 812,
  "completion_tokens": 143,
  "total_tokens": 955,
  "prompt_tokens_details": { "cached_tokens": 640 }
}

Repeating the same system prompt within 5 minutes bills those tokens at the cheaper cached rate. These are exactly the numbers debited from your wallet.

Errors

Status	Code	Meaning
401	invalid_api_key	Key is wrong or was revoked by regeneration.
401	wrong_product_key	You used the other product's key.
402	subscription_expired	No active plan — renew in the console.
402	insufficient_balance	Wallet exhausted — recharge in the console.
403	account_disabled	Account disabled by an administrator.

What this model is — and is not

dhandare-4-mini is a deterministic work model, not a general LLM. It serves defined task shapes — trading-signal narration, post-trade reflection, bull/bear debate, assistant planning (including read-only SQL over a supplied schema) and answer formatting — and returns the same output for the same input, every time, with no GPU and no third-party API behind it. Prompts outside those shapes get an explicit "unsupported task" reply: this model never invents prose to look clever.