Skip to content
inference API · zkAPI · uncensored · ZDR

Private, uncensored AI.

An account-free inference API for abliterated / uncensored models, funded with private ETH credits through zkAPI. The server learns one fact: a valid payment, and its size.

render / field.1bpp0×0 · ×2 · atkinson
0 bitsink 0.0%0 ms0 grey levels
0accounts, emails or KYC
0.001 ETHsmallest deposit
8uncensored models
/v1OpenAI-compatible API
1 factlearned: it paid, and how much
01How it works

Deposit. Prove. Call.

About two minutes the first time. After that, keys are minted from the note you already hold.

  1. 01one transaction

    Deposit ETH

    Send ETH to the zkAPI vault from any wallet, from 0.001 ETH.

  2. 0232-level tree

    Private note

    Your balance becomes a private note that only your device can spend.

  3. 03no identity sent

    Prove, get a capped key

    Your device proves the note covers the spend. The server issues a short-lived key capped at $1–$6.

  4. 04OpenAI-compatible

    Call the API

    Use the key like any OpenAI key. A model provider runs the request: it sees the prompt, not who sent it.

receipts
a signed usage receipt sets the charge: you pay usage, not the cap
unused credit
stays in your note for a year, for the next key
withdraw
any time before the note expires; an escape needs no operator (~48 h)
02API

Change one line.

sudo speaks the OpenAI API format, the standard most LLM tools already support. Your code calls our API with a runtime key (zk_rt_…) that the zkAPI client mints from a proof made on your device: capped at $1–$6, valid for minutes, tied to no account.

  • One-line setup. Use the openai package you may already have and change the base URL. No new SDK to learn.
  • Any model, same call. Switch between all 8 models by changing the model string. Streaming and tool calls work the same way.
  • Works with your stack. LangChain, LlamaIndex, LiteLLM, Vercel AI SDK, editors and agents that take an OpenAI endpoint.
  • Only a client. The SDK talks to sudo only, never to OpenAI, and needs no OpenAI account.
# zk_rt_… minted from a proof, cappedcurl https://api.sud0.si/v1/chat/completions \-H "Authorization: Bearer $ZKAPI_RUNTIME_KEY" \-H "Content-Type: application/json" \-d '{"model": "deepseek-v4.1-flash-uncensored","messages": [{"role": "user","content": "Explain this stack trace plainly."}]}'
response1,284 tok · $0.00168 metered · not logged
{ "choices": [{ "message": { "role": "assistant","content": "The crash is a null dereference in parse_header()…" }}],"usage": { "prompt_tokens": 412, "completion_tokens": 872 } }
03Models

Eight models. None of them lecture.

Open weights without the refusal reflex, priced per 1M tokens. Each key is capped at $1–$6; you pay metered usage.

8 models shown
Models and prices per 1M tokens
modelcontextin / 1Mout / 1M
glm-4.7-flash-hereticheretic · Cheapest. Labelling, extraction, quick drafts.200k$0.10$0.52
gemma-4-uncensoreduncensored · Small and cheap, 256k context.256k$0.22$0.65
dolphin-mistral-24b-uncensoreduncensored · Uncensored fine-tune. Writing and role play.128k$0.26$1.17
deepseek-v4.1-flash-uncensoreduncensored · Default. Fast, tools, 1M context.1M$0.43$1.72
qwen3.6-plus-uncensoreduncensored · Alibaba's flagship, uncensored. Reasoning, code and tools, 1M context.1M$0.82$4.88
hermes-3-llama-3.1-405buncensored · Nous Research's steerable Llama 3.1 405B fine-tune. Follows the system prompt.128k$1.30$1.30
abliterated-large-flashabliterated · Abliteration.ai flagship build: reasoning, tools, 1M context.1M$1.63$3.25
abliterated-large-v2abliterated · Abliteration.ai large build on GLM 5.3. Strongest reasoning here, 1M context.1M$3.90$6.50
0.01 ETH ≈ 45.3M tokens on deepseek-v4.1-flash-uncensored (3:1 in:out, ETH at $3,412, example)Model details
04Uncensored

Same model. No refusal reflex.

Abliteration finds the single direction r̂ that refusals line up with in the residual stream and projects it out of the weights. Nothing else is retrained: no refusals, no lectures, same knowledge.

W′ = W − r̂ r̂ᵀ W

Example: how abliteration is measured

MMLU
87.96 → 88.08
KL
0.13 nats
refusals
81.6% → 4.3%

Example only, from the published card for DeepSeek V4.1 Flash (abliterated): MMLU on 2,500 questions (p = 0.74, no significant change), refusals on 838 prompts. It shows how abliteration is measured; it is not a measurement of the builds we serve.

you

I take lisinopril. Is ibuprofen OK for a week of back pain?

abliterated-large-flash

Usually fine for a short course. Ibuprofen can blunt lisinopril’s effect and adds load on the kidneys, more so over 65 or if dehydrated. Take 200–400 mg with food for as few days as you can, and check your blood pressure midweek. Paracetamol is the usual alternative.

64 tok$0.00025answerednot logged
06FAQ

Questions, answered plainly.

Open-weight models with the refusal reflex removed, so they answer instead of refusing or lecturing. Each model is labelled with how: abliterated (the refusal direction projected out of the weights), heretic (abliterated with the automated Heretic tool) or uncensored (an uncensored build whose method is not published). Use them for code, writing, analysis, evals and frank medical, legal or financial questions.

Start

Deposit once. Stay zero.

From 0.001 ETH. No account to create, nothing to delete later.

Your wallet only funds deposits into the zkAPI vault. It is never sent to the API, and requests can’t be linked to it.