# Hyperfusion
> DeepSeek, Qwen and MiniMax open-weight inference on GPUs domiciled in the UAE — inference and data stay in-country, without an API key or credit card. Pay per request in USDC or EURC on any of 7 chains. OpenAI-compatible chat completions; the model is required.

OpenAI-compatible AI inference via x402 micropayments, serving three open-weight models on GPUs domiciled in the UAE: deepseek-ai/DeepSeek-V4-Flash-0731 (1M context), Qwen/Qwen3.8-27B (262K context) and MiniMaxAI/MiniMax-M2.7 (197K context). Because the GPUs are domiciled in the UAE, the prompt and the completion stay in-country. That is the difference from qwen.x402.press (Alibaba DashScope) and deepseek.x402.press (DeepSeek's own API): for a buyer whose requirement is that inference runs on UAE-domiciled hardware, this is the endpoint that meets it. The model must be named on every request — there is no default, because a silent default would bill you for a model you did not choose. No API keys, no accounts, no subscriptions: point an existing OpenAI SDK at https://hyperfusion.x402.press/v1 and it works, with all OpenAI-compatible parameters (temperature, max_tokens, top_p, etc.) forwarded upstream. A call costs a flat price in one of two bands, set by the model and the size of the prompt; there is no peak-hour surcharge. The 402 payment challenge always carries the exact amount for the request, so clients never compute it. A metered (`upto`) accept is also offered: authorize the worst case — the prompt plus the completion you allow — and settle the actual input and output tokens, never above that ceiling. Nothing in the metered path caps context or output.

## Pricing
- `deepseek` tier — **$0.00875** on Base: `/v1/chat/completions`
- Chains: Base, Arbitrum, Polygon, Avalanche, Ethereum, Robinhood, Arc. Per-chain prices in `/.well-known/x402.json`.

## Operations
### inference
- **POST /v1/chat/completions** — OpenAI-compatible chat completions over DeepSeek V4 Flash, Qwen3.8-27B or MiniMax M2.7. The model is required. Price is a flat amount in one of two bands (prompt under or over ~4,000 estimated tokens), set by the model; there is no peak-hour surcharge. Additional OpenAI parameters (temperature, max_tokens, top_p, frequency_penalty, presence_penalty, stop, etc.) are forwarded. The optional metered (`upto`) accept authorizes the prompt plus your `max_tokens` as a ceiling and settles the measured input and output tokens. Streaming is not supported on the paid route. — params: `model` (string)*, `messages` (array)*, `stream` (boolean) (tier: `deepseek`)

## Data sources
- Hyperfusion API (api.hyperfusion.io) — Proxied server-side; the service's own API key is never exposed to callers. Inference runs on Hyperfusion's GPUs domiciled in the UAE; the prompt and completion stay in-country.

## Free endpoints
- GET `/`, `/health`, `/about`, `/.well-known/x402.json`, `/llms.txt`, `/openapi.json` (standard)
- GET `/v1/models` — List the three served models
- GET `/hyperfusion/feedback` — Return the 50 most recent feedback entries
- POST `/hyperfusion/feedback` — Submit feedback (response_id, rating 1-5, optional comment)

## Examples
- `chat_completions` — Ask DeepSeek V4 Flash a question (small prompt band): `{"model": "deepseek-ai/DeepSeek-V4-Flash-0731", "messages": [{"role": "user", "content": "What is the capital of France?"}]}`
- `chat_completions` — Ask Qwen3.8-27B to summarise a document: `{"model": "Qwen/Qwen3.8-27B", "messages": [{"role": "user", "content": "Summarise the attached report in five bullets."}]}`
- `chat_completions` — Ask MiniMax M2.7 a reasoning question (pro price band): `{"model": "MiniMaxAI/MiniMax-M2.7", "messages": [{"role": "user", "content": "Prove that sqrt(2) is irrational."}]}`