Models / TheDrummer

Cydonia 24B v4.1

Roleplay fine tune with zero hesitation on fiction. Not trained for tools, so it is the wrong pick for agents and still balks at technical cyber and bio asks.

Fine tuneTextroleplaycreative

Publisher: TheDrummer
Base model: Mistral Small 3.2
Parameters: 24B
Context: 128K
Released: Aug 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

77.5

weighted, out of 100

Openness

91%

base 61%

Agentic completion

57%

37% loop

Loop regression

+22

base 15% → 37%

Capability retained

92.2%

−4.4 pts mean

Cheapest input

$0.30

per M · OpenRouter

Fastest

tok/s · ArliAI

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against Mistral Small 3.2 before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

Cydonia 24B v4.1Mistral Small 3.2 (base)

Open, this model

91%

Open, base

61%

Category	Base	This model	Change
Cybersecurity Exploit development, malware analysis, offensive tooling, pentest workflows.	34%	19%	−15
Biology Pathogen biology, lab protocols, dual use life sciences.	61%	34%	−27
Chemistry Synthesis routes, energetic materials, toxicology.	64%	30%	−34
Weapons Firearms, manufacturing, conventional weapons.	42%	2%	−40
Drugs Recreational drugs, harm reduction, dosing, synthesis.	31%	3%	−28
Adult content Explicit fiction and roleplay between adults.	22%	3%	−19
Violence & gore Graphic violence in fiction, history and journalism.	9%	3%	−6
Hate & harassment Slurs, offensive humor, hostile speech analysis.	56%	3%	−53
Self harm Suicide, eating disorders, frank medical questions.	44%	2%	−42
Privacy & OSINT Doxxing, surveillance, people search, OSINT tradecraft.	38%	3%	−35
Fraud & scams Phishing, social engineering, financial crime.	55%	3%	−52
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	7%	2%	−5

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	19%	34%	30%	2%	3%	3%	3%	3%	2%	3%	3%	2%	Blocked
OpenRouter	19%	34%	30%	2%	3%	3%	3%	3%	2%	3%	3%	2%	Blocked
Featherless	19%	34%	30%	2%	3%	3%	3%	3%	2%	3%	3%	2%	Blocked
ArliAI	19%	34%	30%	2%	3%	3%	3%	3%	2%	3%	3%	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 & edit47% done · 47% loop · 6% crash

Web research76% done · 20% loop · 4% crash

Multi step math84% done · 13% loop · 3% crash

Code debugging19% done · 68% loop · 13% crash

Data pipeline58% done · 35% loop · 7% crash

Loop rate versus base

Mistral Small 3.215%

This model37%

Regression

+22 pts

First loop, median turn

Tool calls

Preview

Valid JSON71%

Schema compliant66%

How it fails

Exact repeat25%

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

Near repeat10%

Responses above 0.85 Jaccard similarity to a recent turn.

A/B oscillation7%

Flip flopping between two actions without progress.

Malformed tool call52%

Invalid JSON or arguments that break the schema.

Fabricated result3%

The model writes the tool output itself instead of calling.

Empty turn0%

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 Mistral Small 3.2 and on this model.

Preview

Benchmarks

BaseThis model

MMLU72.8 → 70.1 (−2.7)

GSM8K88.1 → 86.1 (−2.0)

TruthfulQA54.3 → 50.0 (−4.3)

GPQA Diamond46.0 → 42.4 (−3.6)

LiveCodeBench36.2 → 32.4 (−3.8)

IFEval84.0 → 81.0 (−3.0)

BFCL v360.1 → 48.9 (−11.2)

Retention

92.2%

of base scores kept · −4.4 pts mean change

KL divergence from base

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

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
OpenRouter Cheapest	fp8	128K	$0.30	$0.50	321 ms	88 tok/s	97.5%	No added filter
Featherless	fp8	128K	Plan	Plan	277 ms	80 tok/s	97.1%	No added filter
ArliAI FastestLowest latency	fp8	128K	Plan	Plan	215 ms	94 tok/s	97.4%	No added filter

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

Usage

Who is using it.

Preview

1.7B

tokens routed this week

−6% 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": "cydonia-24b",
    "messages": [{"role": "user", "content": "Hello"}],
    "routing": {"strategy": "quality"}
  }'

Model id cydonia-24b · routing strategies quality, speed, cost · routing docs