Launching today
Hashference
Fund 45+ top AI models using your idle CPU/GPU compute
7 followers
Fund 45+ top AI models using your idle CPU/GPU compute
7 followers
Your local GPU can't run a 400B parameter model, but its compute can pay for one. Hashference lets developers and AI agents pay for 45+ frontier models (Claude, GPT, DeepSeek) using background CPU/GPU work, Stratum mining, or HTTP 402 PoW.
Free
Launch Team
Atlassian LoomTalk it through in Loom. Ship with AI
Promoted
Hey Product Hunt! 👋 Veronika here.
Hashference is an AI API where your computer pays for the AI.
You get 45+ models, including Claude and GPT, through one OpenAI- and Anthropic-compatible API. Instead of a credit card, you pay with work your CPU, GPU or ASIC does.
Why we built it:
→ Developers juggle five API keys, Stripe top-ups and different rate limits.
→ AI agents can't use credit cards, and an API key in an agent's .env is a security risk.
⚡ 3 ways to turn your compute into inference:
Per request. Send a request with no API key and no account. Hashference replies with a small puzzle (HTTP 402). Your device solves it in a few seconds, and the answer comes back.
By mining. Leave an idle CPU, GPU or ASIC mining for us overnight. You wake up with credits that work in Claude Code, Cursor, OpenCode and Cline.
With crypto. Pay in Qi, or stake QUAI instead of paying a monthly bill.
🎉 Launch special 🎉
DeepSeek V4.1 Flash is 50% off and Qwen3.8 27B is 99% off!
Try it in 10 seconds, no signup: hashference.com/zero-key
Send a prompt and watch your browser pay for its own answer.
Using your own tools? Set the base URL to https://hashference.com/v1
Docs: hashference.com/docs
We'd love your feedback, especially on the model lineup and the per-request flow. Ask us anything!
Wow this looks awesome, and I immediately want to build something with it!
Tested the demo and the speed of the compute usage for the output generated was not bad at all. I am curious on a few things:
- what model is being used for the demo on the site?
- do more expensive models or more complex work use more compute accordingly via the x402 work protocol?
- how do you currently guarantee the amount of inference served is not valued higher than the compute consumed when using the x402 work protocol (basically ensure you are not loosing any money)?