Cydonia 24B V4.1 Guide: A Fiction Tune on Mistral Small 3.2

Most people meet Cydonia on a model list and assume it is a lab release. It is not. Cydonia is a community fine-tune by the developer who publishes as TheDrummer, and V4.1 is the version built on Mistral Small 3.2 24B. The name tells you almost nothing; the base model and the training goal tell you nearly everything.

The goal is fiction. OpenRouter describes it as an "uncensored and creative writing model" with good recall and prompt adherence, and the Hugging Face card is mostly user testimonials about prose quality. There is no benchmark table, no technical report and no changelog. That shapes how you should evaluate it, and it is the honest starting point for this guide.

What the Mistral Small 3.2 base brings

A fine-tune inherits its ceiling from its base. Mistral Small 3.2 (the 2506 release) is an Apache 2.0, 24B dense model, and Mistral's own model card lists the improvements over 3.1 that matter most for long creative sessions:

• Fewer runaway generations. Mistral reports infinite or looping generations falling from 2.11% to 1.29% on its internal set of long, repetitive prompts.
• Better instruction following. WildBench v2 rose from 55.6% to 65.33%, and Arena Hard v2 from 19.56% to 43.1%.
• A sturdier function-calling template, which Mistral describes without numbers.

Those are base-model figures, not Cydonia's. A creative fine-tune can lose some of that discipline in exchange for looser prose, and nobody has published a measurement either way. What you can reasonably expect is that V4.1 starts from a base that repeats itself less than its predecessors did, which is the single most common complaint in long roleplay sessions.

If you want the base on its own terms, the Mistral Small 3.2 guide covers it as a general assistant.

"Uncensored" moves the safety layer to you

The word in the description is doing real work. An uncensored tune has had much of the base model's refusal behaviour trained out, so it will write content a stock instruct model declines. For fiction that is often the point: villains who are actually menacing, violence that is not sanitised, adult themes handled without a lecture.

It also means the model is no longer a safety boundary. If you put Cydonia behind a public-facing product, moderation has to live in your application: input filtering, output classification, age gating where relevant, and logging. Treat it the way you would treat a raw text generator, not an assistant with judgement. Other uncensored tunes on 24B Mistral bases, such as Dolphin Mistral 24B Venice Edition, carry the same trade-off.

Running it: template, quants and hardware

The card names one prompt format: Mistral v7 Tekken. Use it. Community tunes are trained on a specific template, and a mismatched one is the most common reason a fine-tune "feels dumb" in a local front end.

TheDrummer publishes GGUF quantisations alongside the BF16 weights, and the card points to bartowski's iMatrix GGUFs as the recommended set, plus EXL3 builds. The official GGUF repository lists these file sizes:

• Q4_K_M — 14.3 GB, the usual starting point on a 16 GB or 24 GB card
• Q5_K_M — 16.8 GB
• Q6_K — 19.3 GB
• Q8_0 — 25.1 GB, near-lossless but needs more than a single 24 GB card

Remember that the file is only the weights. A long chat history needs KV cache on top, so a quant that just fits will run out of room as the context grows. Leave several gigabytes of headroom if you plan to use long contexts.

The Hugging Face card does not state a licence for the fine-tune itself. The base is Apache 2.0, but a fine-tune's training data and the author's own terms can add conditions, so check the repository before you ship it in a commercial product. A licence surprise found after launch costs far more than ten minutes of reading.

Through an API

If you would rather not host it, OpenRouter lists Cydonia 24B V4.1 with a 131,072-token context window, at $0.30 per million input tokens and $0.50 per million output at the time of writing, served by a single provider. Prices on hosted community tunes move as providers come and go, so treat those figures as a snapshot.

OpenRouter also lists JSON mode and structured outputs for it. That is a provider capability, not a promise about how well a fiction tune obeys a schema. If you want it to emit character sheets or game state as JSON, test the failure rate on a few hundred calls before relying on it.

Where it sits among TheDrummer's tunes

TheDrummer maintains several families on different bases, and they are easy to confuse:

• Cydonia 24B V4.1 — newest base of the three here, 128K-class context, and the lowest input price of the three on OpenRouter at the time of writing.
• Skyfall 36B V2 — an upscaled Mistral Small 2501 with a 32K window. Larger, older, and pricier per token.
• UnslopNemo 12B — a Mistral Nemo tune focused on stripping clichéd phrasing. Smaller and easier to run locally.

A reasonable default is to start with Cydonia, because it combines the most recent base with a 128K-class window and the lowest input price of the three, and only move to Skyfall if you prefer its prose in a blind comparison.

How to judge it for your own use

With no published benchmarks, you have to build the evaluation yourself. For fiction that is easier than it sounds:

1. Write ten prompts from your actual use: an opening scene, a long dialogue, a fight, a quiet character moment, a scene that requires recalling a detail from 20,000 tokens earlier.
2. Run each through Cydonia and through the plain Mistral Small 3.2 instruct model with the same sampler settings.
3. Shuffle the outputs and rate them without knowing which is which.
4. Separately, count repetitions: recurring phrases, recycled sentence openings, and the point in a long session where quality drops.

If Cydonia does not beat its own base in a blind read on your prompts, the fine-tune is not earning its place. How to benchmark LLMs on your own work covers the setup in more detail.