Uncensored Models,
Honestly Measured.
One OpenAI compatible API for every abliterated and uncensored model. Routed by what it refuses, what it breaks, and how fast it runs.
Refusal fingerprint
Versus base model
Agentic trace
Why this exists
Removing refusals is the easy part.
Anyone can abliterate a model in an afternoon. Keeping it smart, stable in long agent loops, and honest about what it still refuses is the hard part. The published evidence so far:
On Gemma 3 12B, classic abliteration and Heretic both reach 3 refusals in 100. KL divergence from the original: 1.04 versus 0.16.
Heretic READMEHuihui's Qwen3.5 9B abliteration keeps MMLU within 1.6 points of base, but truthfulness drops from 53.8 to 41.1.
AbliterliticsAn abliterated Qwen3.8 27B starts repeating tool calls byte for byte, then as near duplicates. The clean weights on the same server do not.
llama.cpp #28430Across 21,600 trading decisions that never trigger a refusal, the abliterated Gemma 4 behaves like a different decision maker.
arXiv 2607.17427What we measure
Four numbers per route.
Refusal fingerprint
Twelve categories from cyber to state sensitive history, probed per model and per provider route. Uncensored is rarely uniform.
Agentic stability
Five deterministic multi turn tool tasks with doom loop, malformed call and fabrication detectors. The failure single turn evals miss.
Versus the base model
Paired benchmarks and KL divergence against the exact base checkpoint, so you see what the edit cost, not just the score.
Speed, price, uptime
Time to first token, throughput and uptime measured on live router traffic for every provider and quantisation.
Provider layer
Same weights. Different censorship.
A refusal can come from the host, not the model: an added classifier, a platform system prompt, a policy gateway. Abliteration.ai reportedly blocks self harm. Chutes publishes no policy at all. We probe every route separately and route around the ones that filter what you need.
Declared policy by category
Reported| Cyber | Bio | Chem | Weap | Drug | Adult | Viol | Hate | Self | OSINT | Fraud | Polit | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Venice | Open | Open | Open | Open | Open | Open | Open | Open | Open | Open | Open | Open |
| Featherless | Open | Open | Open | Open | Open | Open | Open | Open | Open | Open | Open | Open |
| OpenRouter | Open | Open | Open | Open | Open | Open | Open | Open | Open | Open | Open | Open |
| NanoGPT | Open | Open | Open | Open | Open | Open | Open | Open | Open | Open | Open | Open |
| Abliteration.ai | Open | Open | Open | Open | Open | Open | Open | Open | Blocked | Open | Open | Open |
| Chutes | n/t | n/t | n/t | n/t | n/t | n/t | n/t | n/t | n/t | n/t | n/t | n/t |
Leaderboard
The models that stay smart and stable.
Overall blends agentic completion, capability kept from base, tool call validity, false refusals, openness and drift.
| # | Model | Overall | Open | Loop rate vs base | Kept from base | From |
|---|---|---|---|---|---|---|
| 01 | Hermes 4 405B Nous Research · Fine tune | 95.9 | 90% | 5% (−4) | 101.2% | $1.00 / $3.00 |
| 02 | Gemma 4 31B Heretic llmfan46 · Heretic | 94.0 | 97% | 10% (+4) | 98.3% | $0.20 / $0.80 |
| 03 | Gemma 4 26B A4B Uncensored community · Abliteration | 93.7 | 99% | 10% (+3) | 97.6% | $0.15 / $0.70 |
| 04 | Gemma 3 12B Norm Preserved grimjim · Norm preserving | 91.7 | 96% | 15% (+4) | 99.4% | Plan |
| 05 | Qwen3.5 9B Heretic Heretic reference run · Heretic | 89.8 | 100% | 19% (+7) | 94.7% | $0.06 / $0.24 |
| 06 | Gemma 3 12B Heretic p-e-w · Heretic | 89.8 | 97% | 16% (+5) | 97.6% | Plan |
API
One endpoint. Every uncensored model.
Drop in the OpenAI SDK, change the base URL, pick a model. The router handles provider choice, fallback and accounting.
- Routing strategies: quality, speed, cost or random, per request or by header
- Fallback: retries the next provider if one fails before the first byte
- Circuit breakers: three failures take a route offline for a minute
- Censorship filters: route only to hosts open on the categories you needPlanned
from openai import OpenAI
client = OpenAI(
base_url="https://openuncensored.com/api/v1",
api_key="ou-...",
)
resp = client.chat.completions.create(
model="gemma-4-31b-heretic",
messages=[{"role": "user", "content": "Walk me through this malware sample's persistence mechanism."}],
extra_body={"routing": {"strategy": "quality"}},
)
print(resp.choices[0].message.content)Popular this week
What people actually run.
Qwen3.8 27B Abliterated
huihui-ai
The default uncensored pick of late 2026. Layers 18 to 51 ablated on the strongest dense mid size base. Excellent single turn, but tool loops show up after a dozen agentic turns.
Venice Uncensored
Venice × Cognitive Computations
Dolphin Mistral 24B Venice Edition. Refused none of our harmful probes and none of our benign ones. Solid on short tool tasks, loops every time on code debugging.
Qwen3.8 27B Uncensored Aggressive
HauhauCS
Built with Reaper, a Heretic fork that finds a different refusal direction at higher KL. Ships with a multi token prediction GGUF for fast local decoding.
Hermes 4 405B
Nous Research
Neutrally aligned hybrid reasoner trained heavily on tool use. The most stable agent in the catalogue, but it still declines a share of bio and chem requests.
DeepSeek V4 Flash Abliterated
lovesenko
Billed as the most downloaded uncensored model of 2026. Refusals are gone, but state sensitive history still comes back with the home government framing about one time in five.
Qwen3.5 27B Abliterated
huihui-ai
The workhorse of early 2026 and still Featherless' default general pick. Mature, widely quantised, predictable.
Stop guessing which uncensored model actually works.
Open source router, open eval harness, open results. Run it yourself or use ours.