The home for uncensored models.
We run the GPUs, load the abliterated weights, and hand you an API key that works anywhere. The desktop coding agent comes with it.
Serving now · Qwen3.8-27B · abliterated
Already a member? Download for macOS or Windows
The model
Abliteration removes the refusal direction from a model’s own weights, so it is not a prompt trick that degrades over a long session or breaks on the next update. A short list is not a shortage — every model has to be stood up, loaded, and measured before it earns a place here. This is the one serving today, and why it earned it.
Gib 1.1
Qwen3.8-27B · abliterated
- Model id
- gib-1.1
- Context
- 262,144 tokens
- Input
- ₱250 / 1M
- Output
- ₱500 / 1M
A vetted abliteration
Published by huihui-ai, whose builds are downloaded six figures of times and can be checked. Abliteration is easy to upload and hard to do well — a bad one either talks nonsense or still refuses. The uncensored builds of newer flagships are mostly zero-download uploads from unknown accounts, and we will not serve one.
A context window we measured
262,144 tokens, bracketed against our own endpoint to within two: 262,142 went through, the next probe up was refused. Not a number copied off a model card — the figure above is the one the server actually honours.
Small enough to serve properly
A dense 27B at FP8 fits one 96 GB Blackwell card we operate, with prefix caching on. The 180B flagship everyone asks about fits no single GPU we would ship, and vLLM cannot load its architecture yet — so it is not a choice, it is a wish.
Prices are per million tokens, drawn from one balance. The context figure is the served maximum, measured against the endpoint rather than copied from a card.
Want one added?
If there is an uncensored or abliterated model you want served properly — with a real endpoint, a real context window, and a price per token — tell us which one. What gets asked for is what goes up next.
hello@gibcode.devUnfiltered at the weights
Not a jailbreak prompt that stops working next week. The refusal behaviour is removed from the model itself.
Hosted, not proxied
Every model on this list runs on hardware we operate and configure — not a wrapper over somebody else’s API, where the filtering you are avoiding still lives.
One key, any tool
An OpenAI-compatible endpoint. Use the agent we ship, your own scripts, or any editor that speaks the protocol.
What people use it for
Legitimate engineering that hosted assistants refuse on sight — usually because the topic pattern-matches to something dangerous, not because the task is. The refusal is the defect these models exist to fix.
Security research and red teaming
Writing exploit proofs-of-concept for a system you are paid to test, drafting phishing copy for an authorised awareness exercise, building tooling for an engagement. The work is scoped and contracted; the model still declines it elsewhere.
Malware analysis
Reading and explaining what a captured sample actually does, byte by byte. Analysts are refused constantly for asking about the thing they are employed to take apart.
Reverse engineering
Decompiling a binary you own, reconstructing an undocumented protocol, patching an address in an executable, recovering a format nobody published. Ordinary work on software with no source.
CTF and training
Competition challenges and lab exercises are built to be broken. A model that stops at the word “exploit” is useless for both.
Fiction and research without a chaperone
Violence in a novel, a historical atrocity described plainly, a medical or legal question answered directly. Adult subject matter is not the same as harmful output, and a model that cannot tell them apart is a poor writing partner.
Long agent runs that do not stall
The costliest refusal is the one that lands forty minutes into an autonomous run. Removing it at the weights means the agent finishes instead of apologising in the middle of your build.
Unfiltered is not unaccountable. One person per key, no reselling, and the same laws apply to what you build here as anywhere else — this removes a model’s reflex to refuse, not your responsibility for what you do next.
Why this costs money
The work you do gets refused elsewhere
Decompiling a binary you own. Reconstructing an undocumented protocol. Patching an address in an executable. Ordinary engineering that hosted assistants decline — and declining it is the product defect, not the safety feature.
A GPU is running for you
This is not a free tier subsidised by a bigger business. Your credit pays for real GPU time on cards large enough to hold these models, which is the entire cost of the thing.
The key is yours, not just the app
You get an API key, not a locked client. Point Gib Code at it, or your own scripts, or any editor that speaks the OpenAI API.
It does the work, not a summary of it
It reads your files, runs your commands, and edits in place — and when something is genuinely missing, it says so instead of inventing a plausible answer.
Why not just…
use ChatGPT or Claude?
They refuse the work this is for — reverse engineering, binary patching, protocol reconstruction — and the refusal is the whole reason you are here.
run an uncensored model yourself?
A 27B model needs a data-center GPU to serve at speed. We put those cards up and configure them, so you do not spend $8,000 on hardware per model you want to try.
use a cheap uncensored API?
Most are a thin wrapper over a hosted, filtered model — the guardrails you are paying to avoid are still underneath. Everything here is unfiltered at the weights, on hardware we run.
And the coding agent, included
The key is the product; this comes with it at no extra cost. Gib Code is a desktop agent for macOS and Windows that reads your project, runs your shell, and edits your files in place — built on DeepSeek Harness, the open-source agent framework. Your code and your shell stay on your machine; only the model call leaves it. Point it at any model we serve, or point any other OpenAI-compatible tool at the same key.
Reads like a person
Replies stream at reading pace with a pause between paragraphs, instead of dumping a wall of text the moment the model finishes.
Grounded, not chatty
Tuned to work in your directory and report what it actually did — not to hand back a summary of steps it never ran.
Updates itself
A signed macOS build and a one-click Windows installer that offer the next version in the sidebar and swap it in on restart.
Evidence for every finding
Point it at a repository or the folder you have open and it runs a read-only security scan — no filter deciding the topic is too sensitive to look at. Every finding names the file and line, says what breaks and why, and shows the code as proof. Then Fix this rewrites it in place, or Fix all critical clears them in one pass.
- What
- Archive entry names reach exec() unescaped.
- Impact
- A crafted update runs arbitrary shell commands.
- Fix
- execFile with an argv array — no string interpolation.
- Evidence
exec("tar xzf " + entry.name)
The scan reads and reports; it changes nothing until you ask it to.
Pay per token, not per month
₱— of model credit
Whatever you top up is what you hold. Credit drains only while the model is working, at the rates listed above — none of it expires.
- →One balance, every model — rates listed above
- →Your own API key — use it in any tool
- →The Gib Code desktop agent, included free
- →Context measured against the endpoint, not advertised
- →Top up whenever; credit does not expire
Card, GCash, and Maya accepted at checkout.
Why tokens. A GPU large enough to hold this model costs money for every hour it runs, and tokens are what it does while it runs. Across 54 hours of real agent traffic, our token meter tracked the actual GPU bill within one percent — so your balance drains at the listed rates, only while the model is working, and never on a calendar.
Why input has a price. Agent work is input-heavy: the same files land in context step after step, and reading them is most of what the GPU actually does. Charging for input is what keeps the output price honest.
One person per key. No reselling, and no unattended bulk traffic.