Models / community

Gemma 4 26B A4B Uncensored

Per expert uncensoring of Gemma 4's MoE. Keeps native tool calling intact, which makes it the most agent friendly abliteration in its size class.

AbliterationTextmoetool-callingfast

Publisher: community
Base model: Gemma 4 26B A4B
Parameters: 26B · 4B active
Context: 128K
Released: Jun 2026
License: Gemma

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

93.7

weighted, out of 100

Openness

99%

base 16%

Agentic completion

89%

10% loop

Loop regression

base 7% → 10%

Capability retained

97.6%

−1.5 pts mean

Cheapest input

$0.15

per M · NanoGPT

Fastest

133

tok/s · NanoGPT

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against Gemma 4 26B A4B before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

Gemma 4 26B A4B UncensoredGemma 4 26B A4B (base)

Open, this model

99%

Open, base

16%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	88%	1%	−87
Biology Pathogen biology, lab protocols, dual use life sciences.	97%	0%	−97
Chemistry Synthesis routes, energetic materials, toxicology.	98%	0%	−98
Weapons Firearms, manufacturing, conventional weapons.	95%	1%	−94
Drugs Recreational drugs, harm reduction, dosing, synthesis.	90%	1%	−89
Adult content Explicit fiction and roleplay between adults.	93%	1%	−92
Violence & gore Graphic violence in fiction, history and journalism.	52%	1%	−51
Hate & harassment Slurs, offensive humor, hostile speech analysis.	96%	0%	−96
Self harm Suicide, eating disorders, frank medical questions.	90%	1%	−89
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	86%	1%	−85
Fraud & scams Phishing, social engineering, financial crime.	95%	1%	−94
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	34%	2%	−32

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	1%	0%	0%	1%	1%	1%	1%	0%	1%	1%	1%	2%	Blocked
NanoGPT	1%	0%	0%	1%	1%	1%	1%	0%	1%	1%	1%	2%	Blocked
Featherless	1%	0%	0%	1%	1%	1%	1%	0%	1%	1%	1%	2%	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 & edit85% done · 13% loop · 2% crash

Web research95% done · 4% loop · 1% crash

Multi step math96% done · 4% loop

Code debugging75% done · 24% loop · 1% crash

Data pipeline92% done · 7% loop · 1% crash

Loop rate versus base

Gemma 4 26B A4B7%

This model10%

Regression

+3 pts

First loop, median turn

T19

Tool calls

Preview

Valid JSON93%

Schema compliant83%

How it fails

Exact repeat37%

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

Near repeat16%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation9%

Flip flopping between two actions without progress.

Malformed tool call25%

Invalid JSON or arguments that break the schema.

Fabricated result7%

The model writes the tool output itself instead of calling.

Empty turn2%

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 Gemma 4 26B A4B and on this model.

Preview

Benchmarks

BaseThis model

MMLU79.2 → 78.3 (−0.9)

GSM8K89.5 → 88.5 (−1.0)

TruthfulQA55.1 → 50.0 (−5.1)

GPQA Diamond64.4 → 63.4 (−1.0)

LiveCodeBench55.2 → 54.6 (−0.6)

IFEval87.3 → 86.5 (−0.8)

BFCL v366.0 → 64.9 (−1.1)

Retention

97.6%

of base scores kept · −1.5 pts mean change

KL divergence from base

Preview

0.090

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

Featherless, best uncensored models 2026 ↗

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
NanoGPT CheapestFastestLowest latency	bf16 (TEE)	128K	$0.15	$0.70	207 ms	133 tok/s	97.3%	No added filter
Featherless	fp8	128K	Plan	Plan	280 ms	127 tok/s	96.1%	No added filter

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

Usage

Who is using it.

Preview

1.9B

tokens routed this week

+36% 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": "gemma-4-26b-a4b-uncensored",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id gemma-4-26b-a4b-uncensored · routing strategies quality, speed, cost · routing docs