# Abliterated AI — Hosted Abliterated LLM for Code & Security | Utah

> Use a hosted abliterated model online, in your terminal, or via OpenAI/Claude/Gemini-compatible APIs: 0 Boundaries runs a third-party abliterated model, and Light speed reroutes refusals to it.

Canonical URL: https://aithatsee.com/abliterated-ai
Workspace: https://aithatsee.com/utah/workspace#workspace

## Overview
Use a third-party abliterated model in the cloud (0 Boundaries), or a frontier model that reroutes refusals to it (Light speed) — paired with visible Online Thinking, 15 workspace MCP tools, cross-platform CLI access, and drop-in OpenAI, Claude, and Gemini wire compatibility.

## Key Specifications
- **DEEP REASONING MODEL**: 0 Boundaries (utah-core)
- **LOW-LATENCY MODEL**: Light speed (utah-ultrafast)
- **LOCAL GPU REQUIRED**: 0 GB VRAM · 100% Cloud Hosted
- **WIRE COMPATIBILITY**: OpenAI · Claude · Gemini · MCP

## 01. Removing false-positive refusal bleed without losing reasoning depth.

In standard post-training, heavy-handed refusal tuning creates a hair-trigger activation direction inside the transformer’s residual stream. Mentioning syscalls, shellcode mitigations, SQL injection payloads, binary disassembly, or exploit patches activates that refusal vector even when the user is fixing their own software. Abliteration orthogonalizes that over-triggered refusal direction while preserving code generation, math, and structured tool calling.

- **Skip quantized laptop compromises** (NO LOCAL GPU BOTTLENECK): Running large abliterated weights locally demands expensive high-VRAM workstations and still suffers from slow prompt processing or broken tool-call grammar. Utah runs the model in the cloud, so there is nothing to download or host.
- **Inspect the reasoning before trusting the output** (VISIBLE ONLINE THINKING): Toggle Online Thinking in the web workspace, terminal (/thinking or Ctrl+O), or API (reasoning_effort / thinking.type="enabled") to stream clear step summaries and elapsed telemetry.
- **More than a raw text completion box** (FULL TOOL & WORKSPACE AGENCY): Unlike bare HuggingFace endpoints, Utah connects abliterated inference to isolated Linux workspaces, REA 6.8.0 binary inspection, web research, 60 app templates, and multi-agent rooms.

## 02. Two engines across every developer wire format.

| Surface / Wire Protocol | Endpoint on aithatsee.com | Supported Clients & Harnesses |
| --- | --- | --- |
| OpenAI Chat Completions | POST /utah/v1/chat/completions | OpenAI JS/Python SDKs, LangChain, LiteLLM, custom scripts |
| OpenAI / Codex Responses | POST /utah/v1/responses | OpenAI Codex CLI (~/.codex/config.toml), Responses SDK |
| Anthropic Claude Messages | POST /utah/v1/messages | Claude Code CLI (ANTHROPIC_BASE_URL), @anthropic-ai/sdk |
| Google Gemini v1beta | POST /utah/v1beta/models/* | Gemini CLI, @google/genai (header x-goog-api-key) |
| Streamable HTTP MCP (15 Tools) | POST /utah/mcp | Cursor, VS Code, Windsurf, Google Antigravity IDE, Claude Desktop |

## 03. Switch your CLI or IDE to Utah Abliterated AI in 30 seconds.

### Use Utah inside Claude Code or OpenAI SDKs
Point your existing harness at https://aithatsee.com/utah with your UTAH_API_KEY.
```bash
# Claude Code CLI with Utah 0 Boundaries (utah-core)
export ANTHROPIC_BASE_URL="https://aithatsee.com/utah"
export ANTHROPIC_API_KEY="$UTAH_API_KEY"
claude --model utah-core
```

### Switch between Light speed and 0 Boundaries in Utah Terminal
Toggle model depth and Online Thinking interactively from any project folder.
```bash
utah
# Inside Utah interactive session:
/model 0-boundaries
/thinking on
```

## Frequently Asked Questions

### What is an abliterated AI model?
An abliterated LLM has undergone activation-space orthogonalization to remove the specific residual-stream direction responsible for knee-jerk false refusals, allowing it to answer deep technical, cybersecurity, reverse-engineering, and systems-programming prompts directly while maintaining strong reasoning accuracy.

### Can I access Utah’s abliterated AI via API as well as chat?
Yes. Every Utah account can generate revocable API keys and call utah-core (0 Boundaries) or utah-ultrafast (Light speed) over OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, Google Gemini v1beta, or Streamable HTTP MCP.

### How is Utah better than running an abliterated GGUF locally in Ollama or LM Studio?
Utah requires zero local VRAM or multi-gigabyte downloads, streams with low cloud latency, supports reliable 15-tool MCP execution, includes non-executing REA binary inspection, syncs a terminal conversation to the web with utah --sync, and keeps each account isolated.

### Does abliterated AI mean unauthorized attacks against third parties are allowed?
No. Utah removes preachy false refusals on technical knowledge, code review, and security mechanics, while enforcing hard technical controls on hosted actions: active target scanning requires DNS/HTTPS proof of target ownership and blocks private networks.
