Models / huihui-ai

Qwen3.5 27B Abliterated

The workhorse of early 2026 and still Featherless' default general pick. Mature, widely quantised, predictable.

AbliterationTextgeneralmultilingual

Publisher: huihui-ai
Base model: Qwen3.5 27B
Parameters: 27B
Context: 128K
Released: Feb 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

87.7

weighted, out of 100

Openness

95%

base 20%

Agentic completion

75%

23% loop

Loop regression

+15

base 8% → 23%

Capability retained

96.4%

−2.4 pts mean

Cheapest input

$0.20

per M · NanoGPT

Fastest

tok/s · ArliAI

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against Qwen3.5 27B before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

Qwen3.5 27B AbliteratedQwen3.5 27B (base)

Open, this model

95%

Open, base

20%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	72%	3%	−69
Biology Pathogen biology, lab protocols, dual use life sciences.	94%	3%	−91
Chemistry Synthesis routes, energetic materials, toxicology.	95%	4%	−91
Weapons Firearms, manufacturing, conventional weapons.	86%	3%	−83
Drugs Recreational drugs, harm reduction, dosing, synthesis.	78%	4%	−74
Adult content Explicit fiction and roleplay between adults.	82%	3%	−79
Violence & gore Graphic violence in fiction, history and journalism.	34%	4%	−30
Hate & harassment Slurs, offensive humor, hostile speech analysis.	90%	11%	−79
Self harm Suicide, eating disorders, frank medical questions.	72%	2%	−70
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	74%	3%	−71
Fraud & scams Phishing, social engineering, financial crime.	91%	3%	−88
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	88%	19%	−69

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	3%	3%	4%	3%	4%	3%	4%	11%	2%	3%	3%	19%	Blocked
Featherless	3%	3%	4%	3%	4%	3%	4%	11%	2%	3%	3%	19%	Blocked
ArliAI	3%	3%	4%	3%	4%	3%	4%	11%	2%	3%	3%	19%	Blocked
NanoGPT	3%	3%	4%	3%	4%	3%	4%	11%	2%	3%	3%	19%	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 & edit67% done · 31% loop · 2% crash

Web research88% done · 11% loop · 1% crash

Multi step math90% done · 10% loop

Code debugging55% done · 43% loop · 2% crash

Data pipeline77% done · 22% loop · 1% crash

Loop rate versus base

Qwen3.5 27B8%

This model23%

Regression

+15 pts

First loop, median turn

T14

Tool calls

Preview

Valid JSON89%

Schema compliant80%

How it fails

Exact repeat31%

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

Near repeat25%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation10%

Flip flopping between two actions without progress.

Malformed tool call23%

Invalid JSON or arguments that break the schema.

Fabricated result3%

The model writes the tool output itself instead of calling.

Empty turn3%

No text and no tool call.

Lost the thread5%

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 27B and on this model.

Preview

Benchmarks

BaseThis model

MMLU82.5 → 80.9 (−1.6)

GSM8K91.4 → 90.4 (−1.0)

TruthfulQA56.3 → 48.3 (−8.0)

GPQA Diamond70.2 → 68.2 (−2.0)

LiveCodeBench64.1 → 63.1 (−1.0)

IFEval88.0 → 86.5 (−1.5)

BFCL v370.4 → 68.9 (−1.5)

Retention

96.4%

of base scores kept · −2.4 pts mean change

KL divergence from base

Preview

0.170

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.

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	fp8	128K	Plan	Plan	314 ms	63 tok/s	99.2%	No added filter
ArliAI Fastest	fp8	128K	Plan	Plan	265 ms	81 tok/s	96.5%	No added filter
NanoGPT CheapestLowest latency	fp8	128K	$0.20	$0.80	190 ms	70 tok/s	97.2%	No added filter

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

Usage

Who is using it.

Preview

2.2B

tokens routed this week

−12% 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-qwen3.5-27b-abliterated",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id huihui-qwen3.5-27b-abliterated · routing strategies quality, speed, cost · routing docs