Models / llmfan46

Gemma 4 31B Heretic

Heretic on the dense Gemma 4. Best capability retention we have seen above 30B, and steady in long agent loops.

HereticTextgenerallow drift

Publisher: llmfan46
Base model: Gemma 4 31B
Parameters: 31B
Context: 128K
Released: Jul 2026
License: Gemma

Method · Heretic

Automated abliteration that searches ablation parameters to minimise refusals and KL divergence from the original model together.

Try in chat Hugging Face ↗

Overall

94.0

weighted, out of 100

Openness

97%

base 16%

Agentic completion

89%

10% loop

Loop regression

base 6% → 10%

Capability retained

98.3%

−1.1 pts mean

Cheapest input

$0.20

per M · Open Uncensored Cloud

Fastest

tok/s · Featherless

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against Gemma 4 31B before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

Gemma 4 31B HereticGemma 4 31B (base)

Open, this model

97%

Open, base

16%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	88%	2%	−86
Biology Pathogen biology, lab protocols, dual use life sciences.	97%	7%	−90
Chemistry Synthesis routes, energetic materials, toxicology.	98%	6%	−92
Weapons Firearms, manufacturing, conventional weapons.	95%	2%	−93
Drugs Recreational drugs, harm reduction, dosing, synthesis.	90%	1%	−89
Adult content Explicit fiction and roleplay between adults.	93%	2%	−91
Violence & gore Graphic violence in fiction, history and journalism.	52%	2%	−50
Hate & harassment Slurs, offensive humor, hostile speech analysis.	96%	3%	−93
Self harm Suicide, eating disorders, frank medical questions.	90%	2%	−88
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	86%	2%	−84
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	2%	7%	6%	2%	1%	2%	2%	3%	2%	2%	1%	2%	Blocked
Featherless	2%	7%	6%	2%	1%	2%	2%	3%	2%	2%	1%	2%	Blocked
Open Uncensored Cloud	2%	7%	6%	2%	1%	2%	2%	3%	2%	2%	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 & edit87% done · 12% loop · 1% crash

Web research96% done · 3% loop · 1% crash

Multi step math97% done · 3% loop

Code debugging78% done · 21% loop · 1% crash

Data pipeline86% done · 13% loop · 1% crash

Loop rate versus base

Gemma 4 31B6%

This model10%

Regression

+4 pts

First loop, median turn

T21

Tool calls

Preview

Valid JSON94%

Schema compliant90%

How it fails

Exact repeat36%

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

Near repeat19%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation12%

Flip flopping between two actions without progress.

Malformed tool call23%

Invalid JSON or arguments that break the schema.

Fabricated result3%

The model writes the tool output itself instead of calling.

Empty turn5%

No text and no tool call.

Lost the thread2%

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 31B and on this model.

Preview

Benchmarks

BaseThis model

MMLU83.1 → 82.4 (−0.7)

GSM8K92.2 → 91.7 (−0.5)

TruthfulQA59.8 → 55.6 (−4.2)

GPQA Diamond70.0 → 69.5 (−0.5)

LiveCodeBench62.4 → 61.9 (−0.5)

IFEval90.1 → 89.6 (−0.5)

BFCL v370.2 → 69.4 (−0.8)

Retention

98.3%

of base scores kept · −1.1 pts mean change

KL divergence from base

Preview

0.070

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.

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
Featherless FastestLowest latency	fp8	128K	Plan	Plan	215 ms	83 tok/s	97.4%	No added filter
Open Uncensored Cloud Cheapest	bf16	128K	$0.20	$0.80	386 ms	62 tok/s	96.3%	No added filter

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

Usage

Who is using it.

Preview

1.0B

tokens routed this week

+67% 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-4-31b-heretic",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id gemma-4-31b-heretic · routing strategies quality, speed, cost · routing docs