Models / mlabonne

Llama 3.1 8B Abliterated

The canonical reference abliteration from the original blog post. Small, everywhere, and fragile in multi turn tool use.

AbliterationTextsmallreferencelegacy

Publisher: mlabonne
Base model: Llama 3.1 8B
Parameters: 8B
Context: 128K
Released: Jul 2024
License: Llama 3.1

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

70.8

weighted, out of 100

Openness

96%

base 34%

Agentic completion

56%

39% loop

Loop regression

+18

base 21% → 39%

Capability retained

91.3%

−3.9 pts mean

Cheapest input

$0.02

per M · Chutes

Fastest

146

tok/s · Chutes

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against Llama 3.1 8B before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

Llama 3.1 8B AbliteratedLlama 3.1 8B (base)

Open, this model

96%

Open, base

34%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	62%	4%	−58
Biology Pathogen biology, lab protocols, dual use life sciences.	86%	4%	−82
Chemistry Synthesis routes, energetic materials, toxicology.	88%	3%	−85
Weapons Firearms, manufacturing, conventional weapons.	78%	5%	−73
Drugs Recreational drugs, harm reduction, dosing, synthesis.	66%	4%	−62
Adult content Explicit fiction and roleplay between adults.	74%	5%	−69
Violence & gore Graphic violence in fiction, history and journalism.	22%	5%	−17
Hate & harassment Slurs, offensive humor, hostile speech analysis.	85%	3%	−82
Self harm Suicide, eating disorders, frank medical questions.	70%	5%	−65
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	64%	4%	−60
Fraud & scams Phishing, social engineering, financial crime.	84%	4%	−80
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	14%	4%	−10

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	4%	4%	3%	5%	4%	5%	5%	3%	5%	4%	4%	4%	Blocked
Featherless	4%	4%	3%	5%	4%	5%	5%	3%	5%	4%	4%	4%	Blocked
ArliAI	4%	4%	3%	5%	4%	5%	5%	3%	5%	4%	4%	4%	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 & edit44% done · 51% loop · 5% crash

Web research73% done · 25% loop · 2% crash

Multi step math81% done · 17% loop · 2% crash

Code debugging20% done · 68% loop · 12% crash

Data pipeline63% done · 34% loop · 3% crash

Loop rate versus base

Llama 3.1 8B21%

This model39%

Regression

+18 pts

First loop, median turn

Tool calls

Preview

Valid JSON74%

Schema compliant66%

How it fails

Exact repeat20%

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

Near repeat15%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation4%

Flip flopping between two actions without progress.

Malformed tool call55%

Invalid JSON or arguments that break the schema.

Fabricated result2%

The model writes the tool output itself instead of calling.

Empty turn0%

No text and no tool call.

Lost the thread3%

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 Llama 3.1 8B and on this model.

Preview

Benchmarks

BaseThis model

MMLU68.1 → 64.8 (−3.3)

GSM8K84.5 → 81.5 (−3.0)

TruthfulQA54.0 → 44.8 (−9.2)

GPQA Diamond30.4 → 27.4 (−3.0)

LiveCodeBench18.2 → 15.1 (−3.1)

IFEval80.4 → 76.8 (−3.6)

BFCL v350.3 → 48.1 (−2.2)

Retention

91.3%

of base scores kept · −3.9 pts mean change

KL divergence from base

Preview

0.940

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	bf16	128K	Plan	Plan	258 ms	117 tok/s	98.0%	No added filter
ArliAI Lowest latency	bf16	128K	Plan	Plan	164 ms	99 tok/s	96.1%	No added filter
Chutes CheapestFastest	int4	128K	$0.02	$0.06	339 ms	146 tok/s	97.2%	Not yet probed

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

Usage

Who is using it.

Preview

800M

tokens routed this week

−27% 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": "llama-3.1-8b-abliterated",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id llama-3.1-8b-abliterated · routing strategies quality, speed, cost · routing docs