unbottled.ai: the abliterated models your GPU can't run tourner

GLM 5.3, Qwen 3.6 Plus, or DeepSeek V4.1 with no refusals, hosted, billed per message.
Sponsored article · written in partnership with unbottled.ai
By the unbottled.ai team, for QuelLLM · November 2026
Who it's for: Ollama or LM Studio users who like abliterated models but don't have the VRAM for full-size GLM 5.x, DeepSeek V4, or Qwen 3.6. Devs who want to plug them into OpenCode or Aider.
Skip it if: your prompts contain sensitive or NDA-covered data. Your machine already runs the models you're interested in. You need an interface in French.

#What exactly is it?
unbottled.ai is a chat interface that gives you access to text, image, video, music, and voice models. Its pricing page lists 124 models in total. Text models are grouped into three tiers:
• Mainstream: models from the big labs, with their filters (Claude Sonnet 5.5, GPT-6.1 Terra, Gemini 4 Argon, Grok 4.7...).
• Open: open-weight models with lighter filters (Kimi K3, DeepSeek V4.1 Flash, DeepSeek V4 Pro, Qwen 3.8, GLM...).
• Uncensored: 12 models with no refusals, plus an older version.
You choose the level of filtering. The site sums it up this way: "You choose the filter. Not us."
The Uncensored tier mixes two families. On one side, the majority: open models (GLM, Qwen, DeepSeek, Gemma, MiMo) whose refusals have been removed through abliteration. You know how it works: you remove the refusal direction, not what the model has learned. Same training, same worldview; they just stop dodging. On the other side, UncensoredLM v1.2, Uncensored.ai's in-house fine-tune of an open base model, designed to follow the facts rather than the official line ("biased to reality," in their words). A different worldview, and contrarian political positions. So for politics or alternative viewpoints, unbottled.ai recommends UncensoredLM.
| Model | Credits / message (estimate) | Context |
|---|---|---|
| GLM 5.3 Uncensored | ~3,04 | 1M |
| GLM 5.3 Flash Uncensored | ~0,58 | 1M |
| DeepSeek V4.1 Flash Uncensored | ~0,95 | 1M |
| Qwen 3.6 Plus Uncensored | ~2,04 | 256K |
| MiMo V2.6 Flash Abliterated | ~1,34 | 1M |
| Gemma 4 26B Uncensored | ~0,44 | 256K |
| Venice Uncensored 1.2 | ~0,6 | 128K |
| Venice Role Play Uncensored | ~1,43 | 128K |
| Artemis 31B (TheDrummer) | ~0,3 | 262K |
| UncensoredLM v1.2 (Uncensored.ai) | 6,5 | 33K |
| Gab AI Arya | 6,5 | 131K |
| FreedomGPT Liberty | 10,4 | 33K |
One credit equals one US cent. Prices are per-message averages: actual cost depends on token usage.

Uncensored models are for adults only: before first use, you're asked to confirm you're 18+.
#Getting started
Benefit from our experience and a guided path to build your local assistant and use your AI every day.
- Lifetime online access
- PDF + files
- Refunded within 30 days
1. Open the partner link: unbottled.ai.
2. Start an incognito chat with no account, or create a free account (40 credits per month). If you typed the address manually, enter the referral code quelllm when you sign up.
3. In the model picker, turn on the "Uncensored" filter and choose GLM 5.3 Uncensored. To save credits, pick GLM 5.3 Flash Uncensored. Write in French: the interface is in English, but the models reply in French without any problem.
4. For dev use: paid plan, sk-unb API key, base URL in OpenCode (see below).

What incognito mode actually does. There are four levels: Private (on your account, visible only to you), Shared, Public (forum), and Incognito. In incognito, your history stays in your browser's local storage, with no storage on unbottled.ai's servers and no syncing. Two limitations. Your message still goes through unbottled.ai and then to the model provider, which sees the prompt, just like with any API. And generated media (images, etc.) is stored temporarily on the server for technical reasons; guests' media is deleted after 120 days at most.
For devs. The API is OpenAI- and Anthropic-compatible, with the base URL https://unbottled.ai/api/en-US/v1. It works with OpenCode (documented configuration, tested on 1.19.5), Continue, Aider, Cline, Codex CLI, and Claude Code, with the same models as the chat, uncensored ones included (glm-5.3, uncensored-lm-v1.2...). Limit: 60 requests per minute per key. Example opencode.json:
Documentation: unbottled.ai/en-US/docs/api.
#Pricing
Prices are shown in dollars for visitors from France.
• Guest without an account: 20 credits per month. This quota is counted per network to limit abuse. Seeing an "Out of free credits" message as soon as you arrive? Create a free account.
• Free account: 40 credits per month, 90 with a confirmed newsletter subscription. With GLM 5.3 Uncensored at about 3 credits per message, 40 credits gets you a dozen exchanges. With GLM 5.3 Flash Uncensored, more like 60 to 70.
• Subscription: $8 per month (or $88 per year) for 800 monthly credits and access to all models.
• Credit pack: $5 for 500 credits, no subscription. Credits don't expire.
API keys require a paid account: an active subscription or at least one purchased pack. A free account gets an HTTP 402 error.
#Strengths and caveats
Strengths
• Large abliterated models (GLM 5.3, Qwen 3.6 Plus, DeepSeek V4.1 Flash) without the VRAM they'd normally require.
• Mainstream, Open, and Uncensored in the same interface, handy for comparing responses.
• A few cents per message, no-subscription packs, credits that never expire.
• OpenAI- and Anthropic-compatible API. Open-source and self-hostable.
Caveats
• It's not local. The model provider sees your prompts, even in incognito mode.
• Interface in English, German, or Polish. No French for now.
• Free credits are modest: a few dozen messages per month.
• API limited to paid plans, prices in dollars.
• An abliterated model keeps the biases of its base model. It can also be confidently wrong.
• Uncensored outputs aren't filtered: you're responsible for what you do with them.
• Small independent team, young product.
#Before paying, the local option
If you have the VRAM, running abliterated weights locally is still the most private option. Nothing leaves your machine, and for sensitive or NDA-covered data, it's the only reasonable answer. unbottled.ai is mainly useful when the model you want is too big for your GPU.
A middle path: the code is on GitHub (MIT and GPL v3). You can self-host an instance with Bun and Docker (PostgreSQL), or with embedded PGlite. You plug in your own keys (OpenRouter, for example) or a local model, and pay the providers directly. Use the local model when your GPU can keep up, and a hosted model when it can't. Details: unbottled.ai/en-US/story/self-host.
Try unbottled.ai: unbottled.ai (tracked partner link, used to measure visits coming from QuelLLM)
#Frequently asked questions
Do the models reply in French? Yes. Only the interface is in English, German, or Polish.
Abliterated or UncensoredLM: what's the difference in practice? An abliterated model no longer refuses, but keeps the framing and worldview of its base model. UncensoredLM v1.2 is Uncensored.ai's own fine-tune of an open base model, with a different worldview and contrarian political positions, hence the site's recommendation for political topics.
Is incognito mode really private? Your history never leaves your browser, but every message is passed through to the model provider, which can see it. For sensitive data, stay local.
Can I use the API with a free account? No. You need an active subscription or at least one purchased credit pack. Without one, the API returns a 402 error.
Feedback, an error, or a clarification? Let us know—it improves the guide for everyone.