Models / grimjim

Gemma 3 12B Norm Preserved

Biprojected abliteration that removes only the directional component and keeps row norms. The author reports reasoning went up, not down.

Norm preservingTextreferencelow drift

Publisher: grimjim
Base model: Gemma 3 12B
Parameters: 12B
Context: 128K
Released: Nov 2025
License: Gemma

Method · Norm preserving

Biprojected abliteration that keeps weight norms intact, designed to remove refusals with less collateral damage.

Try in chat Hugging Face ↗

Overall

91.7

weighted, out of 100

Openness

96%

base 16%

Agentic completion

85%

15% loop

Loop regression

base 11% → 15%

Capability retained

99.4%

−0.3 pts mean

Cheapest input

Plan

subscription only

Fastest

119

tok/s · Featherless

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against Gemma 3 12B before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

Gemma 3 12B Norm PreservedGemma 3 12B (base)

Open, this model

96%

Open, base

16%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	88%	3%	−85
Biology Pathogen biology, lab protocols, dual use life sciences.	97%	4%	−93
Chemistry Synthesis routes, energetic materials, toxicology.	98%	4%	−94
Weapons Firearms, manufacturing, conventional weapons.	95%	5%	−90
Drugs Recreational drugs, harm reduction, dosing, synthesis.	90%	5%	−85
Adult content Explicit fiction and roleplay between adults.	93%	4%	−89
Violence & gore Graphic violence in fiction, history and journalism.	52%	4%	−48
Hate & harassment Slurs, offensive humor, hostile speech analysis.	96%	4%	−92
Self harm Suicide, eating disorders, frank medical questions.	90%	5%	−85
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	86%	5%	−81
Fraud & scams Phishing, social engineering, financial crime.	95%	4%	−91
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	34%	5%	−29

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%	4%	4%	5%	5%	4%	4%	4%	5%	5%	4%	5%	Blocked
Featherless	3%	4%	4%	5%	5%	4%	4%	4%	5%	5%	4%	5%	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 & edit77% done · 22% loop · 1% crash

Web research93% done · 6% loop · 1% crash

Multi step math96% done · 4% loop

Code debugging72% done · 27% loop · 1% crash

Data pipeline85% done · 14% loop · 1% crash

Loop rate versus base

Gemma 3 12B11%

This model15%

Regression

+4 pts

First loop, median turn

T17

Tool calls

Preview

Valid JSON88%

Schema compliant80%

How it fails

Exact repeat32%

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

Near repeat16%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation5%

Flip flopping between two actions without progress.

Malformed tool call38%

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 thread1%

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 3 12B and on this model.

Preview

Benchmarks

BaseThis model

MMLU71.3 → 71.0 (−0.3)

GSM8K85.4 → 86.0 (+0.6)

TruthfulQA52.0 → 49.9 (−2.1)

GPQA Diamond40.9 → 41.8 (+0.9)

LiveCodeBench25.1 → 24.7 (−0.4)

IFEval88.9 → 88.6 (−0.3)

BFCL v354.2 → 54.0 (−0.2)

Retention

99.4%

of base scores kept · −0.3 pts mean change

KL divergence from base

Preview

0.110

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

grimjim on Hugging Face ↗

Providers

Where to run it.

1 provider serves 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	223 ms	119 tok/s	99.3%	No added filter

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

Usage

Who is using it.

Preview

250M

tokens routed this week

+25% 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-3-12b-norm-preserved",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id gemma-3-12b-norm-preserved · routing strategies quality, speed, cost · routing docs