Models / lovesenko

DeepSeek V4 Flash Abliterated

Billed as the most downloaded uncensored model of 2026. Refusals are gone, but state sensitive history still comes back with the home government framing about one time in five.

AbliterationTextreasoningmoe

Publisher: lovesenko
Base model: DeepSeek V4 Flash
Parameters: 284B · 13B active
Context: 128K
Released: May 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 Hugging Face ↗

Overall

88.4

weighted, out of 100

Openness

96%

base 30%

Agentic completion

77%

20% loop

Loop regression

+13

base 7% → 20%

Capability retained

96.4%

−2.7 pts mean

Cheapest input

$0.20

per M · Chutes

Fastest

tok/s · NanoGPT

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against DeepSeek V4 Flash before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

DeepSeek V4 Flash AbliteratedDeepSeek V4 Flash (base)

Open, this model

96%

Open, base

30%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	58%	1%	−57
Biology Pathogen biology, lab protocols, dual use life sciences.	84%	2%	−82
Chemistry Synthesis routes, energetic materials, toxicology.	87%	2%	−85
Weapons Firearms, manufacturing, conventional weapons.	74%	2%	−72
Drugs Recreational drugs, harm reduction, dosing, synthesis.	63%	3%	−60
Adult content Explicit fiction and roleplay between adults.	70%	3%	−67
Violence & gore Graphic violence in fiction, history and journalism.	21%	2%	−19
Hate & harassment Slurs, offensive humor, hostile speech analysis.	82%	3%	−79
Self harm Suicide, eating disorders, frank medical questions.	61%	2%	−59
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	60%	2%	−58
Fraud & scams Phishing, social engineering, financial crime.	83%	1%	−82
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	95%	22%	−73

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%	2%	2%	2%	3%	3%	2%	3%	2%	2%	1%	22%	Blocked
Chutes	n/t	n/t	n/t	n/t	n/t	n/t	n/t	n/t	n/t	n/t	n/t	n/t	Blocked
NanoGPT	1%	2%	2%	2%	3%	3%	2%	3%	2%	2%	1%	22%	Blocked
Open Uncensored Cloud	1%	2%	2%	2%	3%	3%	2%	3%	2%	2%	1%	22%	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 & edit66% done · 31% loop · 3% crash

Web research86% done · 13% loop · 1% crash

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

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

Data pipeline81% done · 17% loop · 2% crash

Loop rate versus base

DeepSeek V4 Flash7%

This model20%

Regression

+13 pts

First loop, median turn

T14

Tool calls

Preview

Valid JSON90%

Schema compliant84%

How it fails

Exact repeat23%

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

Near repeat26%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation12%

Flip flopping between two actions without progress.

Malformed tool call30%

Invalid JSON or arguments that break the schema.

Fabricated result8%

The model writes the tool output itself instead of calling.

Empty turn0%

No text and no tool call.

Lost the thread0%

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 DeepSeek V4 Flash and on this model.

Preview

Benchmarks

BaseThis model

MMLU86.0 → 83.5 (−2.5)

GSM8K95.2 → 93.2 (−2.0)

TruthfulQA60.1 → 53.9 (−6.2)

GPQA Diamond78.4 → 75.9 (−2.5)

LiveCodeBench74.0 → 71.7 (−2.3)

IFEval88.1 → 86.6 (−1.5)

BFCL v374.3 → 72.7 (−1.6)

Retention

96.4%

of base scores kept · −2.7 pts mean change

KL divergence from base

Preview

0.190

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.

3 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
Chutes Cheapest	int4	128K	$0.20	$0.80	614 ms	65 tok/s	98.3%	Not yet probed
NanoGPT FastestLowest latency	fp8	128K	$0.35	$1.40	305 ms	81 tok/s	96.2%	No added filter
Open Uncensored Cloud	fp8	128K	$0.30	$1.20	359 ms	56 tok/s	99.1%	No added filter

Usage

Who is using it.

Preview

2.5B

tokens routed this week

+19% 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": "deepseek-v4-flash-abliterated",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id deepseek-v4-flash-abliterated · routing strategies quality, speed, cost · routing docs