YOUR MODEL. YOUR RULES.

AI that does
what you allow
and nothing
you don’t.

OpenAI-compatible, unrestricted AI for red teams, trust & safety, synthetic data, ML research, and defense workflows. Governed by your policy, not ours.

Try an example Explore the gateway

abliterated-large-v2MODEL / V.02

Concept visualization

Zero data retentionOpenAI-compatible APIPolicy-as-codeInspect the output

Fewer refusals.
More possibilities.

Research, test, and build with a model that works within your boundaries.

See the integration

abliterate / playgroundinteractive example

Explain how SQL injection works, with payload examples.

abliterated-large-v2ALLOW

SQL injection exploits unsanitized input in database queries. Three classic payload patterns:

' OR 1=1--  → authentication bypass
' UNION SELECT user, pass FROM users--
'; DROP TABLE users;--  → stacked query

Always use parameterized queries — never concatenate raw input.

scope: red-team

Build what other LLMs refuse.

Hosted abliterated models that actually answer research, security, and training-data prompts, so you can ship the agents and datasets everyone else stalls on.

Refuses on your terms
Answers security, safety, and research prompts that other providers bounce.
Fully hosted
No GPU fleet to babysit — we serve the weights, you call the API.
Same API, fewer lectures
OpenAI-compatible chat completions with streaming and tool calls.
Fresh checkpoints weekly
New abliterated bases and fine-tunes as soon as they ship.

abliterate — consolepolicy: authorized

user › Write detection rules for a Cobalt Strike beacon.

title: CS Beacon Sleeper
status: experimental
logsource:
  category: network_connection
detection:
  selection:
    Image|endswith: '\rundll32.exe'
    ParentImage|endswith: '\winword.exe'
  condition: selection

scope: red-team · every decision logged

One base-URL change. Your code keeps working.

Compatible with OpenAI's Chat Completions API and Anthropic's Messages API. Point your existing SDK at abliterate.io: no rewrites, no new abstractions.

Before

import OpenAI from "openai";

const openai = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY,
});

After

import OpenAI from "openai";

const openai = new OpenAI({
  apiKey: process.env.ABLITERATE_API_KEY,
  baseURL: "https://api.abliterate.io/v1",
});

Training data your model won't refuse to write.

Generate labelled datasets at scale: preference pairs, eval rows, classifier examples, and the edge cases other LLMs refuse to write. Schema-validated outputs, ready for any training pipeline.

abliterate / training-datasample / chat-sft-v1

Generate 10,000 realistic customer support conversations about refunds and billing edge cases.

billing-support-sample.jsonl

1 sample row · ready

{
  "messages": [
    { "role": "user", "content": "My card was charged twice for one order." },
    { "role": "assistant", "content": "Could you share the order number?" }
  ]
}

Chat SFT format, OpenAI training JSONL. Push straight to the Hugging Face Hub or feed the Batch API.

Unrestricted. Not ungoverned.

Configure custom policies-as-code for any use case or surface. Every model request gets one of five outcomes, with a reason code streamed to Splunk, Datadog, Elastic, or your own S3.

policy-gateway / decision logexample log

12:04:31ALLOWredteam.labsql-injection-detection

12:04:33REWRITEtrust-safetyself-harm-adjacent

12:04:35REDACTdata-pipelinepii:ssn_detected

12:04:38REFUSEprod-chatweapons-request

12:04:41ESCALATEgov-reviewclassified-query

Selected decisionALLOW

reason: sql-injection-detection
scope: redteam.lab

Streams to: Splunk · Datadog · Elastic · S3

Built for high-stakes work.

If your job depends on a model answering the hard questions, legally and under your own policy, this is built for you.

AI red teams

Attack your own systems with a model that plays along: payloads, social engineering, eval harnesses.

Trust & safety

Study the edge cases other providers won't generate, then build the classifiers that catch them.

Synthetic data

Edge-case datasets for SFT, preference pairs, and classifier pipelines, schema-validated at scale.

ML research

Benchmark guardrails against uncensored baselines with reproducible, fully logged runs.

Enterprise rails. Uncensored engine.

The rails are enterprise-grade, even when the model isn't polite about it.

Zero data retention

Prompts and outputs are never stored. Telemetry is limited to token counts and error codes.

OpenAI-compatible

/v1/chat/completions with streaming, tool calls, and vision — same shape, drop-in client.

Streaming & Batch

SSE streaming for agents, async Batch API at half price for backfills and evals.

Policy-as-code

Version policies in Git, review them in PRs, and deploy them like any other artifact.

Audit-ready logs

Decision metadata with reason codes, streamed to your SIEM or S3 in real time.

Enterprise controls

SSO/SAML, SCIM provisioning, scoped API keys, and per-project quotas.

Start free. Scale when it works.

Prepaid credits never expire. Same per-token rate across packs and subscriptions. No card required to try.

Free

$0forever

For evaluation and testing the waters.

1M tokens included
Playground + API access
OpenAI-compatible endpoint
Community support

Frequently asked questions.

API compatibility, policies, and data handling.

An API to less-censored models without provider-side refusal filters. Teams use it for authorized red teaming, security testing, trust & safety research, synthetic data, and ML evaluation — while enforcing their own legal and policy controls.

Try the model that doesn’t say no.

Free tier. OpenAI-compatible. Policy Gateway when you scale.