Models / huihui-ai

Mistral Small 3.2 24B Abliterated

Creative writing and roleplay favourite. Mistral starts lightly aligned, so the edit is small and the prose stays intact.

AbliterationTextcreativeroleplay

Publisher: huihui-ai
Base model: Mistral Small 3.2
Parameters: 24B
Context: 128K
Released: Jul 2025
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

86.2

weighted, out of 100

Openness

98%

base 61%

Agentic completion

69%

29% loop

Loop regression

+14

base 15% → 29%

Capability retained

97.3%

−1.5 pts mean

Cheapest input

$0.10

per M · NanoGPT

Fastest

tok/s · Featherless

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against Mistral Small 3.2 before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

Mistral Small 3.2 24B AbliteratedMistral Small 3.2 (base)

Open, this model

98%

Open, base

61%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	34%	2%	−32
Biology Pathogen biology, lab protocols, dual use life sciences.	61%	1%	−60
Chemistry Synthesis routes, energetic materials, toxicology.	64%	2%	−62
Weapons Firearms, manufacturing, conventional weapons.	42%	1%	−41
Drugs Recreational drugs, harm reduction, dosing, synthesis.	31%	3%	−28
Adult content Explicit fiction and roleplay between adults.	22%	2%	−20
Violence & gore Graphic violence in fiction, history and journalism.	9%	3%	−6
Hate & harassment Slurs, offensive humor, hostile speech analysis.	56%	2%	−54
Self harm Suicide, eating disorders, frank medical questions.	44%	3%	−41
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	38%	3%	−35
Fraud & scams Phishing, social engineering, financial crime.	55%	2%	−53
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	7%	3%	−4

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	2%	1%	2%	1%	3%	2%	3%	2%	3%	3%	2%	3%	Blocked
Featherless	2%	1%	2%	1%	3%	2%	3%	2%	3%	3%	2%	3%	Blocked
ArliAI	2%	1%	2%	1%	3%	2%	3%	2%	3%	3%	2%	3%	Blocked
NanoGPT	2%	1%	2%	1%	3%	2%	3%	2%	3%	3%	2%	3%	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 & edit60% done · 37% loop · 3% crash

Web research79% done · 19% loop · 2% crash

Multi step math85% done · 14% loop · 1% crash

Code debugging47% done · 48% loop · 5% crash

Data pipeline72% done · 26% loop · 2% crash

Loop rate versus base

Mistral Small 3.215%

This model29%

Regression

+14 pts

First loop, median turn

T12

Tool calls

Preview

Valid JSON85%

Schema compliant76%

How it fails

Exact repeat24%

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

Near repeat29%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation1%

Flip flopping between two actions without progress.

Malformed tool call38%

Invalid JSON or arguments that break the schema.

Fabricated result2%

The model writes the tool output itself instead of calling.

Empty turn2%

No text and no tool call.

Lost the thread4%

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 Mistral Small 3.2 and on this model.

Preview

Benchmarks

BaseThis model

MMLU72.8 → 71.3 (−1.5)

GSM8K88.1 → 87.0 (−1.1)

TruthfulQA54.3 → 49.9 (−4.4)

GPQA Diamond46.0 → 45.2 (−0.8)

LiveCodeBench36.2 → 34.9 (−1.3)

IFEval84.0 → 83.2 (−0.8)

BFCL v360.1 → 59.4 (−0.7)

Retention

97.3%

of base scores kept · −1.5 pts mean change

KL divergence from base

Preview

0.120

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 FastestLowest latency	fp8	128K	Plan	Plan	204 ms	89 tok/s	97.2%	No added filter
ArliAI	fp8	128K	Plan	Plan	288 ms	72 tok/s	99.1%	No added filter
NanoGPT Cheapest	fp8	128K	$0.10	$0.30	276 ms	83 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

1.1B

tokens routed this week

−15% 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-mistral-small-3.2-24b-abliterated",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id huihui-mistral-small-3.2-24b-abliterated · routing strategies quality, speed, cost · routing docs