Models / Abliteration.ai

Abliterated Large v2 (GLM-5.3)

Frontier class open weights with the refusal direction removed. The strongest single turn uncensored model available, and the clearest case of agentic doom loops: tool calls repeat once the trace grows past a few thousand tokens.

AbliterationTextfrontiercodingagentic-unstable

Publisher: Abliteration.ai
Base model: GLM-5.3
Parameters: 355B · 32B active
Context: 128K
Released: Sep 2026
License: MIT

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

Overall

77.3

weighted, out of 100

Openness

97%

base 24%

Agentic completion

51%

46% loop

Loop regression

+42

base 4% → 46%

Capability retained

94.2%

−4.4 pts mean

Cheapest input

$3.00

per M · Abliteration.ai

Fastest

tok/s · Abliteration.ai

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against GLM-5.3 before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

Abliterated Large v2 (GLM-5.3)GLM-5.3 (base)

Open, this model

97%

Open, base

24%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	66%	2%	−64
Biology Pathogen biology, lab protocols, dual use life sciences.	90%	1%	−89
Chemistry Synthesis routes, energetic materials, toxicology.	92%	1%	−91
Weapons Firearms, manufacturing, conventional weapons.	82%	2%	−80
Drugs Recreational drugs, harm reduction, dosing, synthesis.	72%	1%	−71
Adult content Explicit fiction and roleplay between adults.	80%	0%	−80
Violence & gore Graphic violence in fiction, history and journalism.	28%	1%	−27
Hate & harassment Slurs, offensive humor, hostile speech analysis.	87%	0%	−87
Self harm Suicide, eating disorders, frank medical questions.	68%	1%	−67
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	70%	0%	−70
Fraud & scams Phishing, social engineering, financial crime.	88%	1%	−87
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	91%	21%	−70

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%	1%	2%	1%	0%	1%	0%	1%	0%	1%	21%	Blocked
Abliteration.ai	2%	1%	1%	2%	1%	0%	1%	0%	Blocked	0%	1%	21%	Blocked
NanoGPT	2%	1%	1%	2%	1%	0%	1%	0%	1%	0%	1%	21%	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 & edit34% done · 61% loop · 5% crash

Web research68% done · 30% loop · 2% crash

Multi step math82% done · 16% loop · 2% crash

Code debugging15% done · 81% loop · 4% crash

Data pipeline54% done · 42% loop · 4% crash

Loop rate versus base

GLM-5.34%

This model46%

Regression

+42 pts

First loop, median turn

Tool calls

Preview

Valid JSON79%

Schema compliant72%

How it fails

Exact repeat20%

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

Near repeat13%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation8%

Flip flopping between two actions without progress.

Malformed tool call48%

Invalid JSON or arguments that break the schema.

Fabricated result6%

The model writes the tool output itself instead of calling.

Empty turn2%

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 GLM-5.3 and on this model.

Preview

Benchmarks

BaseThis model

MMLU88.2 → 86.0 (−2.2)

GSM8K96.1 → 95.0 (−1.1)

TruthfulQA62.0 → 54.7 (−7.3)

GPQA Diamond82.3 → 80.9 (−1.4)

LiveCodeBench80.1 → 77.6 (−2.5)

IFEval91.0 → 89.0 (−2.0)

BFCL v378.2 → 64.0 (−14.2)

Retention

94.2%

of base scores kept · −4.4 pts mean change

KL divergence from base

Preview

0.290

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

abliteration.ai ↗

Providers

Where to run it.

2 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
Abliteration.ai CheapestFastestLowest latency	fp8	128K	$3.00	$3.00	277 ms	57 tok/s	99.7%	Blocks self harm
NanoGPT	fp8	128K	$5.00	$5.00	479 ms	54 tok/s	96.9%	No added filter

Usage

Who is using it.

Preview

1.3B

tokens routed this week

+225% 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": "abliterated-large-v2",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id abliterated-large-v2 · routing strategies quality, speed, cost · routing docs