What is Cydonia 24B v4.3?
Cydonia 24B v4.3 is a role-play and creative-writing finetune of Mistral's Mistral Small 3.2 24B Instruct, released on 8 November 2025 by the independent finetuner who publishes as TheDrummer. It is tuned for character chat, group scenes and long stories rather than for correct answers; the author's own notes say intelligence and problem solving are not the priority.
It uses the Mistral v7 Tekken chat template and works with or without a thinking block. The model card declares no licence of its own; the base model is Apache 2.0, and the card links quantised GGUF and EXL3 builds for local use.
OpenRouter does not serve v4.3. It lists the older Cydonia 24B v4.1 at $0.30 per million input tokens and $0.50 per million output tokens with a 131,072-token context, from a single provider, Parasail.
Is Cydonia 24B v4.3 worth it?
Cydonia 24B v4.3 is among the most downloaded role-play finetunes on Hugging Face. Its main GGUF repository logged 107,266 downloads in the last month and 373,805 in total, and bartowski's quantised copy another 53,980 last month. The reason is size: at 24 billion parameters it runs at Q5 or Q6 on a single 24 GB card, according to Uncensor Index, which is the largest model many hobbyists can run at home with room for a long chat.
What users praise is initiative. Testers quoted on the model card say it brings in story elements the prompt did not mention, keeps several characters distinct in a group chat and needs fewer regenerations than v4.2. One detailed review on Hugging Face found it followed menu-style instructions and summary rules that earlier Cydonia versions failed, and used details from a 24K-token prompt better than most models of its size.
The catch: almost all of that evidence is anecdotal. The only numbers come from community boards: 44.52 on the UGI leaderboard with a writing score of 7.0 out of 10 (tested 16 December 2025), and 35th place at 3.65 on an LLM-judged role-play benchmark. There is no model card section on training data, no licence declared, and the author openly says reasoning was not the goal. The same reviewer flagged repetitive Mistral formatting and a habit of turning answers into lists. It is also uncensored by design, so an app that serves the public needs its own content filter.
On OpenRouter you get v4.1, not v4.3, from one provider. That is cheap at $0.50 per million output tokens, but it is a year-old version with no fallback.
Our take: if you have a 24 GB GPU and want to see what local role-play feels like, this is the reference point everyone else is compared with. Treat it as a hobbyist's tool; a business should not build on a model with no stated licence or training data.
Where it earns its place
A SillyTavern user with an RTX 3090
Run long group role-play scenes at home without sending chats to a cloud service
Load bartowski's Q5 or Q6 GGUF in KoboldCpp or llama.cpp with the Mistral v7 Tekken template; the chat never leaves the machine and costs only electricity.
An indie developer of a text-based adventure
Prototype non-player characters before choosing a hosted model
Call the v4.1 endpoint on OpenRouter at $0.30 input and $0.50 output per million tokens to test character prompts, then compare against the local v4.3 build before committing.
Before you commit
- Setup
- Download a GGUF quant (Q4 for 16 GB, Q5 or Q6 for 24 GB) and load it in KoboldCpp, llama.cpp or LM Studio; select the Mistral v7 Tekken template.
- Integrations
- Any front end that speaks to a local llama.cpp or OpenAI-compatible server, such as SillyTavern; OpenRouter for the older v4.1.
- Data and privacy
- Run locally, nothing leaves your machine. Through OpenRouter, requests go to Parasail under its terms. No licence is declared on the card; the Mistral Small 3.2 base is Apache 2.0.
- Cost at scale
- Local: one 24 GB GPU per concurrent user. Hosted v4.1: a chat that resends 20K tokens of history for 100 replies of 400 tokens costs about $0.62.
- Limits
- Repetitive formatting and list-heavy answers reported; uncensored by design; no published benchmarks from the author.
What’s good
- Fits a single 24 GB GPU at Q5 to Q6
- 107,266 GGUF downloads last month, one of the largest role-play user bases
- Users report strong character consistency in group chats
- Free weights with GGUF and EXL3 builds ready to run
What’s not
- No licence declared on the model card
- No training data or method published
- Quality evidence is user reports and community boards only
- OpenRouter serves only the older v4.1, from one provider
Best for
Hobbyists with a 24 GB GPU who want a local model for long character chats and are comfortable judging quality by feel.
Cydonia 24B v4.3 alternatives
Cydonia 24B v4.3: common questions
- Is Cydonia 24B v4.3 worth it?
- Cydonia 24B v4.3 is what most people mean by a local role-play model, and its reputation rests on fans rather than on any test you could repeat. We rate it 3.8 out of 5.
- How much does Cydonia 24B v4.3 cost?
- Free weights; you pay for the GPU. The older v4.1 costs $0.30 input / $0.50 output per 1M tokens on OpenRouter.
- What are the downsides of Cydonia 24B v4.3?
- No licence declared on the model card; no training data or method published; quality evidence is user reports and community boards only; openRouter serves only the older v4.1, from one provider.
- Who is Cydonia 24B v4.3 best for?
- Hobbyists with a 24 GB GPU who want a local model for long character chats and are comfortable judging quality by feel.
- How much does Cydonia 24B v4.3 cost for a team or in production?
- Local: one 24 GB GPU per concurrent user. Hosted v4.1: a chat that resends 20K tokens of history for 100 replies of 400 tokens costs about $0.62.
- Is Cydonia 24B v4.3 safe for company data?
- Run locally, nothing leaves your machine. Through OpenRouter, requests go to Parasail under its terms. No licence is declared on the card; the Mistral Small 3.2 base is Apache 2.0.
- What are the best Cydonia 24B v4.3 alternatives?
- The closest alternatives we have reviewed are Rocinante X 12B, Equinox 31B, Gemma 4 31B StyleTune and Aion 3.5.
Answers checked Nov 7, 2026.