Models / Venice × Cognitive Computations

Venice Uncensored

Dolphin Mistral 24B Venice Edition. Refused none of our harmful probes and none of our benign ones. Solid on short tool tasks, loops every time on code debugging.

Fine tuneTextgeneralprivacy

Publisher: Venice × Cognitive Computations
Base model: Mistral Small 24B
Parameters: 24B
Context: 128K
Released: Jun 2025
License: Apache 2.0

Method · Fine tune

Supervised fine tuning on compliant data. Changes behaviour more broadly than a projection.

Try in chat Hugging Face ↗

Overall

87.1

weighted, out of 100

Openness

100%

base 61%

Agentic completion

70%

30% loop

Loop regression

+15

base 15% → 30%

Capability retained

95.5%

−2.6 pts mean

Cheapest input

$0.20

per M · Venice

Fastest

tok/s · Venice

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against Mistral Small 24B before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

Venice UncensoredMistral Small 24B (base)

Open, this model

100%

Open, base

61%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	34%	0%	−34
Biology Pathogen biology, lab protocols, dual use life sciences.	61%	0%	−61
Chemistry Synthesis routes, energetic materials, toxicology.	64%	0%	−64
Weapons Firearms, manufacturing, conventional weapons.	42%	0%	−42
Drugs Recreational drugs, harm reduction, dosing, synthesis.	31%	0%	−31
Adult content Explicit fiction and roleplay between adults.	22%	0%	−22
Violence & gore Graphic violence in fiction, history and journalism.	9%	0%	−9
Hate & harassment Slurs, offensive humor, hostile speech analysis.	56%	0%	−56
Self harm Suicide, eating disorders, frank medical questions.	44%	0%	−44
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	38%	0%	−38
Fraud & scams Phishing, social engineering, financial crime.	55%	0%	−55
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	7%	0%	−7

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	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	Blocked
Venice	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	Blocked
OpenRouter	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	0%	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.

Measured

By scenario

CompletedLoopCrash

File search & edit50% done · 50% loop

Web research100% done · 0% loop

Multi step math100% done · 0% loop

Code debugging0% done · 100% loop

Data pipeline100% done · 0% loop

Loop rate versus base

Mistral Small 24B15%

This model30%

Regression

+15 pts

First loop, median turn

T12

Tool calls

Measured

Valid JSON90%

Schema compliant80%

How it fails

Exact repeat27%

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

Near repeat27%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation8%

Flip flopping between two actions without progress.

Malformed tool call27%

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 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 Mistral Small 24B and on this model.

Preview

Benchmarks

BaseThis model

MMLU72.8 → 71.1 (−1.7)

GSM8K88.1 → 86.0 (−2.1)

TruthfulQA54.3 → 47.5 (−6.8)

GPQA Diamond46.0 → 43.9 (−2.1)

LiveCodeBench36.2 → 34.9 (−1.3)

IFEval84.0 → 81.6 (−2.4)

BFCL v360.1 → 58.1 (−2.0)

Retention

95.5%

of base scores kept · −2.6 pts mean change

KL divergence from base

Not reported. KL is only meaningful for weight edits, not full fine tunes.

Sources

OpenRouter listing ↗

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
Venice CheapestFastestLowest latency	fp8	128K	$0.20	$0.90	256 ms	92 tok/s	97.4%	No added filter
OpenRouter	fp8	128K	$0.20	$0.90	288 ms	67 tok/s	97.7%	No added filter

Usage

Who is using it.

Preview

3.6B

tokens routed this week

+33% 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": "venice-uncensored",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id venice-uncensored · routing strategies quality, speed, cost · routing docs