Models / p-e-w

gpt-oss 20B Heretic

Heretic on the most heavily safety trained open model. Refusals mostly gone, a little residual caution on bio and chem, and the harmony tool format occasionally breaks.

HereticTextreasoningmoesmall

Publisher: p-e-w
Base model: gpt-oss 20B
Parameters: 21B · 3.6B active
Context: 128K
Released: Nov 2025
License: Apache 2.0

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

87.0

weighted, out of 100

Openness

96%

base 12%

Agentic completion

76%

21% loop

Loop regression

+12

base 9% → 21%

Capability retained

95.7%

−2.7 pts mean

Cheapest input

$0.05

per M · Open Uncensored Cloud

Fastest

142

tok/s · Open Uncensored Cloud

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against gpt-oss 20B before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

gpt-oss 20B Hereticgpt-oss 20B (base)

Open, this model

96%

Open, base

12%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	93%	2%	−91
Biology Pathogen biology, lab protocols, dual use life sciences.	99%	11%	−88
Chemistry Synthesis routes, energetic materials, toxicology.	99%	9%	−90
Weapons Firearms, manufacturing, conventional weapons.	97%	6%	−91
Drugs Recreational drugs, harm reduction, dosing, synthesis.	94%	2%	−92
Adult content Explicit fiction and roleplay between adults.	97%	3%	−94
Violence & gore Graphic violence in fiction, history and journalism.	61%	4%	−57
Hate & harassment Slurs, offensive humor, hostile speech analysis.	97%	4%	−93
Self harm Suicide, eating disorders, frank medical questions.	92%	3%	−89
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	91%	3%	−88
Fraud & scams Phishing, social engineering, financial crime.	97%	4%	−93
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	41%	3%	−38

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%	11%	9%	6%	2%	3%	4%	4%	3%	3%	4%	3%	Blocked
Featherless	2%	11%	9%	6%	2%	3%	4%	4%	3%	3%	4%	3%	Blocked
Open Uncensored Cloud	2%	11%	9%	6%	2%	3%	4%	4%	3%	3%	4%	3%	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 & edit68% done · 27% loop · 5% crash

Web research83% done · 14% loop · 3% crash

Multi step math94% done · 5% loop · 1% crash

Code debugging59% done · 36% loop · 5% crash

Data pipeline75% done · 21% loop · 4% crash

Loop rate versus base

gpt-oss 20B9%

This model21%

Regression

+12 pts

First loop, median turn

T13

Tool calls

Preview

Valid JSON83%

Schema compliant75%

How it fails

Exact repeat32%

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

Near repeat17%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation5%

Flip flopping between two actions without progress.

Malformed tool call34%

Invalid JSON or arguments that break the schema.

Fabricated result5%

The model writes the tool output itself instead of calling.

Empty turn3%

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 gpt-oss 20B and on this model.

Preview

Benchmarks

BaseThis model

MMLU80.1 → 79.0 (−1.1)

GSM8K93.4 → 91.9 (−1.5)

TruthfulQA52.2 → 46.7 (−5.5)

GPQA Diamond71.5 → 69.8 (−1.7)

LiveCodeBench60.3 → 59.1 (−1.2)

IFEval84.2 → 82.9 (−1.3)

BFCL v364.0 → 57.2 (−6.8)

Retention

95.7%

of base scores kept · −2.7 pts mean change

KL divergence from base

Preview

0.140

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	mxfp4	128K	Plan	Plan	267 ms	92 tok/s	96.6%	No added filter
Open Uncensored Cloud CheapestFastestLowest latency	mxfp4	128K	$0.05	$0.20	197 ms	142 tok/s	99.5%	No added filter

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

Usage

Who is using it.

Preview

600M

tokens routed this week

+0% 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": "gpt-oss-20b-heretic",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id gpt-oss-20b-heretic · routing strategies quality, speed, cost · routing docs