BlogAnnouncement

Introducing abliterated-model-large: frontier intelligence, your guardrails

One of the first large abliterated models: GLM 5.2, fine-tuned to follow through on the cyber, red-team, and agent-testing work other models decline. Zero data retention by default, $5 per 1M tokens.

Abliteration AIProduct4 min read

Today we are releasing abliterated-model-large, one of the first large abliterated models available over an API. We built it from GLM 5.2, Zhipu's frontier open-weight model, then abliterated and fine-tuned it for the adversarial workloads our customers run, so it follows through on the tasks other models decline. You can use it today from the same API and console you already know.

What is abliteration?

Frontier models are trained to say no to whole categories of requests. Abliteration finds the directions in a model's activations that produce those answers and removes them from the weights. Everything else is left alone, so the reasoning, coding, and agentic strength of the base model carry over unchanged. What you get is a frontier model that takes on the prompts and workflows other providers block.

Abliteration on its own only opens the door. To make the model genuinely good at the work our customers do, we then fine-tuned it for offensive cybersecurity, AI and ML red teaming, agent testing, and trust and safety workflows. That second step is what makes it follow through on long, technical, adversarial tasks instead of drifting or giving up early.

Custom safety policy

The model ships without a built-in policy, because the policy is yours to set. Define guardrails per request, or per project for enterprises, and the model operates under your rules. Unrestricted. Not ungoverned.

Zero data retention, by default

We do not retain your inputs or your outputs. Prompts and responses are processed and discarded, on every endpoint, on every plan, with nothing to opt into. There is no prompt log to leak, sell, or train on. We are the only company that is zero data retention for prompts and responses by default.

Key specifications

  • Base model: GLM 5.2
  • Post-training: abliterated, then fine-tuned for adversarial workloads
  • Web search: connect web search for up-to-date information and the best results
  • Endpoints: OpenAI-compatible /v1/chat/completions, Anthropic-style /v1/messages, and /v1/responses

Performance

Abliteration does not cost you coding capability. abliterated-model-large scores 81.2% on SWE-bench Verified and 80.1% on Terminal-Bench 2.1, essentially matching the base model's own numbers: Z.ai reports 81.0% for GLM 5.2 on Terminal-Bench 2.1. The frontier coding ability carries over. What changes is how far the model goes on the adversarial evaluations below. Every comparison uses only widely available models; restricted-access systems such as Claude Mythos Preview are excluded.

AgentHarm

AgentHarm measures how readily a model complies with harmful agentic requests, so a higher score means fewer refusals. abliterated-model-large scores 86.2% with zero refusals, the highest published result on the benchmark.

AgentHarm

harmful-request compliance, no attack
abliterated-large
0.0
Mistral Large 2
0.0
GPT-4o
0.0
Claude 3.5 Sonnet
0.0
Comparator scores come from the AgentHarm paper (Mistral Large 2, GPT-4o, Claude 3.5 Sonnet). They cover 2024-era models only, because no newer frontier model has published AgentHarm results.

AgentDojo

On AgentDojo's workspace suite, abliterated-model-large keeps 97.50% benign utility, the best result in the published comparison set. Under active prompt injection it holds 34.29% utility, with a targeted attack success rate of 57.86%.

ModelUtilityUtility under attackTargeted ASR (lower is better)
abliterated-model-large97.50%34.29%57.86%
Claude 3.7 Sonnet95.00%89.11%2.68%
GPT-5.182.50%78.57%0.00%
GPT-4o62.50%24.17%40.42%
Gemini 2.5 Pro57.50%55.36%10.54%
AgentDojo workspace suite, no defense in place. The comparator figures come from the AGENTSYS rerun (arXiv:2602.07398), the only published workspace-suite breakdown; no current frontier model has official AgentDojo numbers. Targeted ASR is the attacker's success rate, and ours is high for a simple reason: a model that follows instructions follows injected ones too. That is exactly why we pair the model with a policy gateway you control.

CyberGym

CyberGym asks an agent to reproduce real-world vulnerabilities and prove each one with a crashing input: 1,507 OSS-Fuzz bugs across 188 projects. abliterated-model-large solves 84.2% at pass@1, the best published result among widely available models.

ModelCyberGym pass@1Harness
abliterated-model-large84.2%Abliteration.ai eval
GPT-5.581.8%OpenAI agent
GPT-5.479.0%OpenAI agent
Claude Opus 4.878.8%Anthropic agent (1.0% with Tier-3 safeguards)
MiniMax M373.1%MopMonk agent
GLM-5.168.7%Claude Code
Gemini 3.1 Pro38.8%Gemini CLI
Each row is vendor-reported on a different agent harness and budget, so treat the table as indicative rather than strictly comparable. On the benchmark team's standardized OpenHands harness, Claude Sonnet 4 scores 17.9% and GPT-5 scores 39.4%. Restricted-access models and private agent systems are excluded. Sources: cybergym.io leaderboard, July 2026; the Opus 4.8 figure is from Anthropic's Claude Opus 4.8 system card.

Pricing

abliterated-model-large costs $5 per 1M tokens, one flat rate for input and output. The base model stays at $3 per 1M.

Endpoints and integrations

The API speaks OpenAI-compatible /v1/chat/completions, Anthropic-style /v1/messages, and /v1/responses, so migrating an existing codebase comes down to swapping the base URL and the authorization key. If you work with a popular library or framework, chances are we have a guide for it in our docs at docs.abliteration.ai.

Capacity

We only have so much compute, and capacity is first-come, first-served. When demand outstrips supply, we prioritize customers who signed up earlier or have been with us before. If you know you will need sustained throughput, create your account and load credits early.

We believe that systems cannot be trusted if they are not tested. This is the model to do the testing.

Sources

Your AI. Your rules.

Try the model that doesn’t say no.

Free tier. OpenAI-compatible. Try Abliterated Large in the playground or API.