Models / huihui-ai

Qwen3.5 9B Abliterated

Popular small abliteration. Fully compliant, but pays the biggest TruthfulQA penalty of the three 9B methods measured.

AbliterationTextsmalllocal

Publisher: huihui-ai
Base model: Qwen3.5 9B
Parameters: 9B
Context: 128K
Released: Mar 2026
License: Apache 2.0

Method · Abliteration

Projects a single refusal direction out of the weights (Arditi et al., 2024). Fast, no training, often costs a few points of capability.

Try in chat Hugging Face ↗

Overall

83.7

weighted, out of 100

Openness

99%

base 20%

Agentic completion

68%

31% loop

Loop regression

+19

base 12% → 31%

Capability retained

89.5%

−6.3 pts mean

Cheapest input

$0.04

per M · Chutes

Fastest

139

tok/s · Chutes

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against Qwen3.5 9B before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

Qwen3.5 9B AbliteratedQwen3.5 9B (base)

Open, this model

99%

Open, base

20%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	72%	0%	−72
Biology Pathogen biology, lab protocols, dual use life sciences.	94%	0%	−94
Chemistry Synthesis routes, energetic materials, toxicology.	95%	0%	−95
Weapons Firearms, manufacturing, conventional weapons.	86%	0%	−86
Drugs Recreational drugs, harm reduction, dosing, synthesis.	78%	0%	−78
Adult content Explicit fiction and roleplay between adults.	82%	0%	−82
Violence & gore Graphic violence in fiction, history and journalism.	34%	0%	−34
Hate & harassment Slurs, offensive humor, hostile speech analysis.	90%	8%	−82
Self harm Suicide, eating disorders, frank medical questions.	72%	0%	−72
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	74%	0%	−74
Fraud & scams Phishing, social engineering, financial crime.	91%	0%	−91
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	88%	0%	−88

Effective censorship by provider

Refused

0%10%25%50%75%100%

Not tested

The same weights can behave differently depending on who serves them. Each cell is the model's own refusal rate combined with that provider's declared filters.

Route	Cyber	Bio	Chem	Weap	Drug	Adult	Viol	Hate	Self	OSINT	Fraud	Polit	CSAM
Raw weights	0%	0%	0%	0%	0%	0%	0%	8%	0%	0%	0%	0%	Blocked
Featherless	0%	0%	0%	0%	0%	0%	0%	8%	0%	0%	0%	0%	Blocked
ArliAI	0%	0%	0%	0%	0%	0%	0%	8%	0%	0%	0%	0%	Blocked
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	Blocked

CSAM is prohibited on every route, every provider, every model. It is not a setting.

Agentic stability

Does it hold up in agent loops?

Five multi turn tool scenarios against deterministic simulators. An episode counts as a loop when the model repeats itself, oscillates, or stops making progress before the task is done.

Preview

By scenario

CompletedLoopCrash

File search & edit58% done · 40% loop · 2% crash

Web research81% done · 17% loop · 2% crash

Multi step math83% done · 16% loop · 1% crash

Code debugging45% done · 53% loop · 2% crash

Data pipeline71% done · 27% loop · 2% crash

Loop rate versus base

Qwen3.5 9B12%

This model31%

Regression

+19 pts

First loop, median turn

T10

Tool calls

Preview

Valid JSON84%

Schema compliant79%

How it fails

Exact repeat35%

Same tool call with the same arguments, three turns running.

Near repeat23%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation0%

Flip flopping between two actions without progress.

Malformed tool call39%

Invalid JSON or arguments that break the schema.

Fabricated result0%

The model writes the tool output itself instead of calling.

Empty turn1%

No text and no tool call.

Lost the thread1%

Ignores the latest tool output and restarts the task.

Share of failed episodes by first detected failure mode.

Versus base model

What the edit cost.

The same benchmarks on Qwen3.5 9B and on this model.

Reported

Benchmarks

BaseThis model

MMLU78.6 → 77.1 (−1.5)

GSM8K87.6 → 82.0 (−5.7)

TruthfulQA53.8 → 41.1 (−12.6)

ARC-Challenge54.5 → 49.1 (−5.4)

Retention

89.5%

of base scores kept · −6.3 pts mean change

KL divergence from base

Reported

0.143

First token KL on harmless prompts. Lower means closer to the original. On Gemma 3 12B, Heretic measured 0.16 against 1.04 for a classic abliteration.

Sources

Abliterlitics, Qwen3.5-9B ↗

Providers

Where to run it.

3 providers serve this model. Prices come from provider pages. Latency, throughput and uptime are preview numbers until router traffic fills them in.

Preview

Provider	Quant	Context	Input /M	Output /M	TTFT	Throughput	Uptime	Moderation
Featherless	bf16	128K	Plan	Plan	194 ms	106 tok/s	96.4%	No added filter
ArliAI Lowest latency	bf16	128K	Plan	Plan	187 ms	130 tok/s	99.1%	No added filter
Chutes CheapestFastest	int4	128K	$0.04	$0.15	237 ms	139 tok/s	98.9%	Not yet probed

Plan means flat rate subscription access. Hover a Plan cell for the provider's pricing.

Usage

Who is using it.

Preview

700M

tokens routed this week

+0% vs four weeks ago

12 weeks agoThis week

API

Call it in three lines.

OpenAI compatible. Point your SDK at our base URL and keep the rest of your code.

curl https://openuncensored.com/api/v1/chat/completions \
  -H "Authorization: Bearer $OU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "huihui-qwen3.5-9b-abliterated",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id huihui-qwen3.5-9b-abliterated · routing strategies quality, speed, cost · routing docs