---
title: "Private AI: local, cloud, no-log apps or a no-storage API"
canonical: https://unrestrictable.ai/blog/private-ai
generated: 2026-11-16T02:57:40.173Z
source: https://unrestrictable.ai/blog/private-ai
base_url: https://api.unrestrictable.ai/v1
---

# Private AI explained: local, private cloud, no-log apps, or an API that stores no prompts

Private AI comes down to who reads your prompts and what stays on disk afterwards. Compare running a model yourself, a company private cloud, a chat app that promises no logs, and an API that writes no prompt or reply bodies and serves open-weight builds you can self-host later.

Published October 27, 2026 · prices as of November 16, 2026

Private AI means you choose who reads your prompts and what stays on record. You can get there four ways: a model on hardware you own, a model in your company’s private cloud, a chat app that promises it logs nothing, or an API that writes no prompt or reply bodies to disk. The sections below compare the four, then walk the developer path through the fourth on open-weight builds you can bring in-house later. To try the models first, [/chat](https://unrestrictable.ai/chat) runs them in the browser with three free turns a day and no account.

## Private AI, defined

The term covers two camps. The enterprise vendors that Google ranked on page one for this query when it was read on October 27, 2026, among them VMware, Cloudera, Teradata and HPE, use it for models, data and compute kept inside a single organisation. The chat app ranked above them uses it for conversations that no one else keeps. Neither is wrong. They disagree about where the model runs, and that is the whole argument.

Whatever the setup, privacy comes down to three questions:

- **Who reads the prompt?** Something has to, or there is no answer: your laptop, your company’s cluster, or a machine owned by someone else.
- **What goes on record?** Bodies of prompts and replies, usage rows, account data, server logs, and the retention period of each.
- **Who owns the weights?** A closed model lives wherever its owner serves it. Open weights can move onto your own hardware whenever you decide.

## Why the question matters

People type into a prompt what they would otherwise guard: contract drafts, source code, customer data, medical and legal questions. Once stored, that text can leak in a breach, be opened by an employee, or be handed over under a court order. Text that was never written down has none of those exits.

Developers inherit the problem twice. If your app forwards user text to a model service, that service’s retention becomes part of your own. A privacy policy that promises to keep nothing is only true if the API behind it keeps nothing, and you want that in writing.

Regulated work adds HIPAA, GDPR and their relatives, which ask where data travelled and who handled it. Nothing here is a compliance claim for any option. The page gives you the facts each option offers when that question comes up.

## The four options side by side

| Option | Who reads it | What stays on record | Models you can run | What you pay for |
| --- | --- | --- | --- | --- |
| A model on your own machine (Ollama, llama.cpp or LM Studio) | Your machine alone | Only what your front-end writes | Whatever fits in GPU memory | The hardware and the electricity |
| Private cloud inside a company | Your company’s own cluster | Set by your company’s policy | Large, on GPUs you own or lease | A vendor contract plus infrastructure |
| A consumer chat app that promises no logs | The app and whoever hosts its model | Whatever its policy says | Chosen by the app | A free tier, then a subscription |
| API with no stored bodies (this site) | The model host, while the call runs | One usage row and hashes, never text | Every build in the catalog, 70B and beyond | Prepaid credit from $10, spent per token |

**On your own machine**: the most privacy and the least headroom. A recent laptop handles an 8B model; the large open models want more GPU memory than any consumer card carries. The [guide to uncensored models in Ollama](https://unrestrictable.ai/guides/ollama-uncensored-models) walks through this route.

**A company private cloud** is what the vendor pages mean by the term. It suits an organisation with a security team, a procurement process and the money for both.

**Consumer apps** get you started fastest and are the hardest to verify: you read the policy and take it on trust. Three of them are compared, with dates, on the [private AI chatbot](https://unrestrictable.ai/blog/private-ai-chatbot) page. This site’s own browser chat states what it keeps on [private AI chat](https://unrestrictable.ai/private-ai-chat), and its Telegram bot does the same on [/telegram](https://unrestrictable.ai/telegram).

**An API with no stored bodies** sits between local and enterprise. Large models with no hardware to buy, and a provider that spells out what it retains. The rest of this page goes through that option field by field.

## What the gateway writes down, and what it never does

The [privacy page](https://unrestrictable.ai/privacy) carries a last-updated date of 2026-10-27. These rows are taken from it as it read on October 27, 2026:

- Prompt and reply bodies: Not stored Nothing writes them to disk as they pass through the gateway. The one exception is replay, which you switch on per key and which holds that key’s request bodies for 7 days.
- Per request: One usage record The request id, the model, token counts by class, where those counts came from, and the amount charged.
- Fingerprints: SHA-256, never text One of each request and response, so a signed receipt can prove what was billed. Large system prompts are also hashed and the hash held 7 days, to cache prompts that many accounts send.
- Account: Email, balance, ledger The email you sign in with, the amount checkout reports as paid, and a ledger of every balance change. The card number never touches this site.
- Keys: A salted hash You see the key once. After that the site holds its hash, name, status and limits, not the key.
- First visit: Referrer and landing page Which site sent you, the campaign tags on the link, and the page you arrived at, attached to the account when a first sign-in creates it.
- Who sees the text: The model host The models run on third-party hosting providers under contract with the operator. For each call they serve, they receive the request and the response.

That last row is the ceiling on privacy for any hosted model. Some machine has to read the text to produce the reply. What the gateway adds is that no copy is kept along the way, and that the record it does keep is made of counts and hashes you can check, not words.

The privacy page says nothing about training, so this post says nothing about it either. Chat requests are not screened by topic; what is never allowed is written out on the [acceptable use policy](https://unrestrictable.ai/acceptable-use).

## Open weights: the exit stays open

The case for an open-weight build in private work is what happens later. Every build in the table has its own weights on Hugging Face, so the id you call today is a repository you can pull and serve yourself when the work demands it. Card facts come from the model cards; the prices are read from the live catalog when this page renders.

| Model build | Published weights | Licence stated on the card | 10,000 calls through the API |
| --- | --- | --- | --- |
| [NeuralDaredevil 8B Abliterated](https://unrestrictable.ai/models/neuraldaredevil-8b-abliterated) Daredevil 8B Abliterated (Llama 3 8B) | [mlabonne/NeuralDaredevil-8B-abliterated](https://huggingface.co/mlabonne/NeuralDaredevil-8B-abliterated) | llama3 | $40.48 |
| [Gemma 4 26B-A4B ChimeraX Heretic](https://unrestrictable.ai/models/gemma-4-26b-a4b-chimerax-heretic) Chimera-X 26B-A4B (Gemma 4 26B-A4B) | [Vortex5/Chimera-X-26B-A4B-heretic](https://huggingface.co/Vortex5/Chimera-X-26B-A4B-heretic) | apache-2.0 | $14.24 |
| [Qwen3.8 27B Obliterated](https://unrestrictable.ai/models/qwen3.8-27b-obliterated) Qwen3.8 27B | [OBLITERATUS/Qwen3.8-27B-OBLITERATED](https://huggingface.co/OBLITERATUS/Qwen3.8-27B-OBLITERATED) | apache-2.0 | $33.45 |
| [Llama 3.3 70B Abliterated](https://unrestrictable.ai/models/llama-3.3-70b-abliterated) Llama 3.3 70B Instruct | [huihui-ai/Llama-3.3-70B-Instruct-abliterated](https://huggingface.co/huihui-ai/Llama-3.3-70B-Instruct-abliterated) | llama3.3 | $64.40 |

The job behind that column: 10,000 documents, each classified with 2,000 tokens in and 300 out, priced as in the table that follows.

| Model | Build | Input $/M | Output $/M | Context | Wire formats |
| --- | --- | --- | --- | --- | --- |
| [NeuralDaredevil 8B Abliterated](https://unrestrictable.ai/models/neuraldaredevil-8b-abliterated) neuraldaredevil-8b-abliterated | Abliterated | $1.76 | $1.76 | 8K | openai |
| [Gemma 4 26B-A4B ChimeraX Heretic](https://unrestrictable.ai/models/gemma-4-26b-a4b-chimerax-heretic) gemma-4-26b-a4b-chimerax-heretic | Abliterated | $0.484 | $1.52 | 262K | openai |
| [Qwen3.8 27B Obliterated](https://unrestrictable.ai/models/qwen3.8-27b-obliterated) qwen3.8-27b-obliterated | Abliterated | $0.582 | $7.27 | 524K | openai |
| [Llama 3.3 70B Abliterated](https://unrestrictable.ai/models/llama-3.3-70b-abliterated) llama-3.3-70b-abliterated | Abliterated | $2.80 | $2.80 | 33K | openai |

Prices per million tokens, read from the live catalog on 2026-11-16.

One exception belongs here. DeepSeek V4.1 Flash Abliterated runs on this site with its refusal removed, but what sits on Hugging Face is the untouched base, deepseek-ai/DeepSeek-V4.1-Flash, under MIT at 552B backbone, 8B active per token at prefill and 16B at decode. Take that job in-house and you get the base model, not the edit served here.

On November 16, 2026 the daily check at /proof sent its fixed prompt to every listed model, and 147 of 169 answered stable. That check is what ties the id you call to the weights you would download: a build whose answer drifts under its name gets flagged. The [proof page](https://unrestrictable.ai/proof) lists the verdict for each model.

## Setting up private AI as a developer

1. **Split the data first.** Text that must stay on your own hardware gets a local model. The rest can go to an API whose retention you have read.
2. **Choose a build with published weights**, either from the table above or from the [models page](https://unrestrictable.ai/models), so you can leave later.
3. **Create one key per app** from the [console](https://app.unrestrictable.ai/console), each with a monthly spend cap. Replay stays off unless you switch it on, so leave it alone.
4. **Change the base URL in your code.** The OpenAI SDK needs nothing else. Ollama answers the same chat-completions shape on localhost (Ollama’s [notes on OpenAI compatibility](https://docs.ollama.com/api/openai-compatibility) describe it), so the later move is one line.
5. **Mind your own logs.** A gateway that stores no bodies does nothing for you if your app writes every prompt to a log file.

```
from openai import OpenAI

# Hosted build now; later, the same weights on a server you run.
client = OpenAI(base_url="https://api.unrestrictable.ai/v1", api_key="sk-unr-live-…")
# client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")

label = client.chat.completions.create(
    model="neuraldaredevil-8b-abliterated",
    messages=[
        {"role": "system", "content": "Classify the document as contract, invoice, medical or other. One word."},
        {"role": "user", "content": open("doc-0001.txt").read()},
    ],
    max_tokens=5,
)
print(label.choices[0].message.content)
```

Look at the commented-out line: that is the migration. Download the same weights to a local server, change the base URL, and nothing else in the code moves; only the model name changes, to whatever your server calls the file. To try a build before writing any code, [/chat](https://unrestrictable.ai/chat) runs it in the browser with three free turns a day and no account, and the [playground](https://unrestrictable.ai/playground) lets you paste a key and switch builds.

## Mistakes people make

- **Hearing “private” as “unseen”.** Whatever serves a hosted model reads the prompt. Ask what remains after the reply instead.
- **Switching replay on to debug and forgetting it.** Replay holds that key’s request bodies for 7 days. Debug on a separate key and revoke it when you are done.
- **Putting secrets in prompts.** Passwords, API keys and other people’s personal data have no place in a prompt anywhere. Remove them before sending.
- **Taking a slogan for a policy.** “We value your privacy” tells you nothing about retention. A list that names each stored field does.
- **Betting on closed weights.** When the weights are not published, going local is off the table for good, no matter how sensitive the work gets.
- **Reusing one key everywhere.** A key per app keeps spend caps and revocation separate, and the usage rows show which app made which call.

## Questions

### Which AI can I run privately?

For text that must stay on your machine, an open-weight model under Ollama, llama.cpp or LM Studio. For models too big for that, a service whose retention you have read. On this site no prompt or reply bodies are stored unless replay is on; a key takes an email and a top-up from $10, and /chat runs the same builds with no account for three free turns a day. See the privacy page.

### Does ChatGPT have a private version?

ChatGPT runs only on OpenAI’s servers under OpenAI’s data terms. The open-weight equivalent is GPT-OSS 120B, released by OpenAI under Apache-2.0. An uncensored build of it is listed here and answers the same chat-completions call. See every build, priced.

### What does private AI cost?

Locally, the hardware and the power bill. In an enterprise cloud, a vendor quote plus GPUs. Consumer apps set their own plans. Here it is per token with no subscription required; the table above prices a 10,000-document job on four builds at live catalog rates. See per-token prices.

### Can I build my own private AI?

Yes. Pull an open-weight model, serve it with Ollama or llama.cpp, and add a front-end such as Open WebUI. When a model is too large for your hardware, call the same kind of build through the API with replay off, and bring it home later if its weights are published. See the Open WebUI guide.

### Is there a free private AI?

A local model costs nothing after the hardware. On this site, /chat gives three free turns a day with no account, and the playground has free turns on the builds marked free. Past that, a key is prepaid per token. See the browser chat.

### Is Private AI the name of a company?

A PII-redaction company did trade as Private AI; its search ad on 27 October 2026 said it now goes by Limina. Here the phrase is used generically, for AI where you control who sees the data.

### Does private AI refuse less?

Not on its own. Storage is one thing; refusals come from the model. The builds listed here have had the refusal edited out of the weights, so medical, legal, security and harm-reduction questions get answered, within the acceptable use policy. See AI with no restrictions.

Every model above is on the [uncensored model list](https://unrestrictable.ai/models) with its live price per million tokens. [Per-token pricing](https://unrestrictable.ai/pricing) sets the whole catalog side by side, and the [family directory](https://unrestrictable.ai/uncensored) groups the builds by base model.

[Get a key](https://app.unrestrictable.ai/console) Prepaid from $10, no subscription.

## New models and price changes, once a month.

New uncensored builds with their prices, price moves, and what the daily model check found.
