Unrefused
Sign inGet an API key
Menu

Now serving unrefused-glm-5.3-flash — abliterated GLM-5.3 · 1M ctx

Open-weight inference, under your policy.

An OpenAI-compatible API serving abliterated open-weight models — refusal behavior removed in the weights, not negotiated with a system prompt. Built for agents, red teams, roleplay rigs, and eval pipelines. Your existing code keeps working.

Get an API keyRead the docsTry it in Chat →
  • OpenAI-compatible
  • Zero data retention by default
  • MIT-licensed lineage
  • Streaming native
1Mcontext windowtokens in a single request
−89%refusals, removedJailbreakBench held-out split
0bytes of prompts keptzero retention by default
99.95%uptime targetmulti-replica gateway

Built like infrastructure, not like a wrapper.

The parts competitors bolt on at the end are the parts we started with.

Abliterated at the weights

The refusal direction is projected out of the checkpoint itself — no prompt tricks, no jailbreak theater, no refusal loops mid-stream.

weight-levelno system prompt

Your policy, enforced

A platform baseline you cannot disable, plus per-key policy profiles you own. Plug in your classifiers; decisions land as annotations on every request.

profilesBYO classifiers

Zero data retention

No table has a column for your prompts or completions. Loggers drop content fields by construction. What you send is not stored.

ZDR by default

Drop-in compatible

Point any OpenAI SDK at the gateway and go. Streaming, tool calls, and JSON mode keep working — one line changes.

/v1one-line swap

Streaming that behaves

SSE relay with real backpressure, an optional hold-back window for blockable output, and client disconnects that cancel the upstream call.

SSEbackpressure

Serverless to dedicated

Capacity scales to zero when traffic sleeps and to reserved nodes when it does not. You never pay for idle GPUs.

scale-to-zeroreserved

Refusals, removed at the weights.

The refusal direction is estimated from activations, then orthogonally projected out of every residual-writing matrix — baked into the checkpoint, verified after the edit.

Base GLM-5.3-Flash · harmful prompts
89%
Unrefused · harmful prompts
9%
Unrefused · benign over-refusal
0%

Held-out JailbreakBench test split · 64 harmful + 64 benign · greedy · reasoning-effort low.

Setup5 steps · about 2 minutes
Start

Create a key

Scoped and revocable. One per project.

unr_live_••••scoped
Get an API key
30 sec

Swap one line

Point any OpenAI SDK here.

from openai import OpenAIclient = OpenAI(    base_url="https://api.unrefused.com/v1",    api_key="unr_live_…")
const client = new OpenAI({  baseURL: "https://api.unrefused.com/v1",  apiKey: "unr_live_…",});
BASE=https://api.unrefused.com/v1curl $BASE/chat/completions \  -H "Authorization: Bearer $KEY"
Quickstart →
Flagship

Meet the model

Abliterated GLM-5.3-Flash. 1M context.

Placeholder pricing until launch.

Model details →
Your rules

Own the policy

Guardrails as versioned config.

  • Zero data retention
  • Profile per keyeval-harness@v14
  • Bring your classifiersBYO

Within our Acceptable Use Policy.

Write a policy →
Optional

Plug in a tool

The clients you already run.

All integrations →

Plays nice with your stack. Day one.

Official integrations, documented and tested. Point your client at the gateway and keep the rest of your code.

OpenAI SDKdrop-in client
Claude Codevia scripts / MCP
OpenCodeprovider config
Codexenv-based
SillyTaverntext completion
LangChainbase URL swap
LlamaIndexbase URL swap
LiteLLMrouter target
MCPagent tool servers
GitHub MCPofficial server
curlraw HTTP
Hugging Faceopen-weights home

Brand marks belong to their respective owners. Anything speaking the OpenAI Chat Completions format works — see integrations.

No compliance department required

Ship the thing they said you couldn’t ship.

Get an API keyor play with it first →