Cydonia, Anubis, Magidonia and Skyfall all start life on The Drummer Hugging Face profile: Mistral and Llama bases fine-tuned for roleplay with the alignment trained out. 10 of those releases run on this site, in /chat and by API. This post covers 5 of them, quoting each card on what the tune is for, pulling prices from the live catalog, and placing the daily check verdict beside every model. The weights and the credit belong to TheDrummer; each card is linked.
The Drummer, the Hugging Face account
TheDrummer (display name Drummer) is a Hugging Face account that puts out fine-tunes of open-weight models for roleplay and creative writing. When we read the profile on 27 October 2026 it showed 222 models, 4,068 followers, one dataset (AmoralQA-v2) and a single external link to a Patreon. Recent activity on the profile includes repos under the BeaverAI organization as well.
Families visible that day: Cydonia and Magidonia at 24B, Skyfall at 31B and 36B, Rocinante and Rivermind between 12B and 16B, Artemis at 31B, Orion at 26B, Behemoth at 123B and 128B, and Anubis at 70B. The cards follow one pattern: base model, chat template, a link to GGUF files with bartowski's iMatrix quants suggested, and a run of user quotes. Sampler values are mostly absent.
10 TheDrummer releases run on this site, in the browser at /chat and by API. Below, 5 of them: the card's own account of what each tune does, today's per-token price, and the daily check on whether the model behind the id is still the one named. The remaining builds each have a model page: Anubis 70B v1, Artemis v1.1, Cydonia 24B v2, Rocinante 12B v1.1, UnslopNemo 12B v4.1.
The five builds, repo by repo
One row per repo on the TheDrummer account, checked on 27 October 2026. Base and template come straight from the card. The downloads figure is the card's "last month" number on that date, so it is already out of date.
| Build | Base model (card) | Chat template | Downloads, last 30 days | Verdict today |
|---|---|---|---|---|
cydonia-24b-v4.3 | Mistral Small 3.2 24B Instruct 2506 | Mistral v7 Tekken | 2,183 | Steady since Oct 23 |
cydonia-24b-v4.1 | Mistral Small 3.2 24B (tagged mistral) | Mistral v7 Tekken | 3,737 | Steady since Oct 26 |
magidonia-24b-v4.3 | Magistral Small 2509 | Mistral v7 Tekken | 941 | Changed Nov 14 |
skyfall-36b-v2 | Mistral Small 2501, upscaled to 36B | Mistral v7 Tekken (Alpaca for adventure) | 3,620 | Steady since Oct 23 |
anubis-70b-v1.1 | Llama 3.3 70B Instruct | Llama 3 | 175 | Changed Nov 15 |
Not served here: Behemoth, Rocinante, Rivermind, Orion and the 31B Skyfall. Their weights are on the profile if you want to run them yourself.
What the cards say each one does
- Cydonia 24B v4.3. The newest Cydonia. Its card promises "much better roleplay" than v4.1 and notes that thinking tags are supported but optional.
- Cydonia 24B v4.1. The release before v4.3, described on its card as "Cydonia Evolved again". Prose comes first; a reviewer quoted on the card rates it "probably the best Cydonia".
- Magidonia 24B v4.3. Cydonia v4.3 rebuilt on the Magistral reasoning base. Per the card, writing style is the main difference between the two.
- Skyfall 36B v2. An upscale with extra training "for creativity and RP". The card points you to Tekken and warns off Metharme.
- Anubis 70B v1.1. Tuned "for character adherence and unalignment", in the card's words. The largest of the five, meant for big casts and personas that must hold.
When people search "the drummer cydonia" they usually mean the 24B line. v4.3 is current, and the Cydonia 24B v4.3 page has the full write-up. v4.1 remains on the list because the reviewers on its card like its prose better, and it is priced differently. Magidonia applies the same tune to Magistral, so a scene that wants a thinking pass before the reply belongs there. Skyfall is the middle size, and that month only one of the five was downloaded more. Anubis, at 70B, exists so characters stay themselves.
No license appears on any of the five cards; the catalog carries whatever the base model's license is. For cydonia-24b-v4.3 that is: none listed (base: Apache-2.0).
Templates, thinking tags, and choosing one
The cards say a lot about templates and little about samplers. Four of the five call for Mistral v7 Tekken, and Anubis calls for Llama 3. Here the template is applied server-side from the model id, so your client sends ordinary messages and never a Tekken tag. The thinking guidance still applies: both v4.3 cards, Cydonia and Magidonia, say a <thinking> block works and that skipping it works too. Skyfall's card says Alpaca fits choose-your-own-adventure play and that Metharme should be avoided. Mag Mell, a merge from another author on the same roleplay table here, is one of the few tunes on this site whose card gives tested sampler values; the Drummer cards leave temperature to you. In /chat it is a slider in Settings: begin at 1.0 and bring it down if the replies wander.
Choosing by the cards and the price rows: Cydonia v4.3 is the current, proven 24B. v4.1 is for anyone who values the reviewers' prose preference over the price. Magidonia is for scenes that gain from a reasoning step. Skyfall is for a larger model. Anubis is for a big cast where one character has to keep its voice across a long chat. The 100-message column below puts a number on each.
Prices today, and what 100 messages cost
Per-million-token rates pulled from the catalog on 2026-11-16. None of these numbers is typed; the table follows the catalog.
| Wire formats | |||||
|---|---|---|---|---|---|
| Cydonia 24B v4.3 | Roleplay | $0.484 | $0.60 | 131K | |
| Cydonia 24B v4.1 | Roleplay | $1.20 | $2.00 | 131K | |
| Magidonia 24B v4.3 | Roleplay | $0.3535 | $0.4235 | 33K | |
| Skyfall 36B v2 | Roleplay | $1.925 | $2.80 | 33K | |
| Anubis 70B v1.1 | Roleplay | $1.085 | $1.085 | 32K |
| Build | Context | What 100 messages cost |
|---|---|---|
| Cydonia 24B v4.3 | 131,072 tokens | $0.11 |
| Cydonia 24B v4.1 | 131,072 tokens | $0.30 |
| Magidonia 24B v4.3 | 32,768 tokens | $0.08 |
| Skyfall 36B v2 | 32,768 tokens | $0.47 |
| Anubis 70B v1.1 | 32,000 tokens | $0.25 |
You top up from $10 by card, WeChat Pay or crypto; plans are optional and credit lasts 12 months. One balance pays for these five and everything else on the models page, and the pricing page lists every rate.
Chat in the browser, or call the API
The quickest route is /chat: three turns a day with no account, a character picker, and the system prompt and temperature under Settings. The roleplay page and uncensored AI roleplay cover that side.
By API, every build accepts the OpenAI chat format at https://api.unrestrictable.ai/v1. Keys come from the console, begin with sk-unr-, and are displayed once. Because the template is added on the server, you send messages rather than Tekken tags. Credit for the key is bought in the console from $10, with no ID check, and the same balance pays for images and video at /media.
curl https://api.unrestrictable.ai/v1/chat/completions \ -H "Authorization: Bearer sk-unr-live-…" \ -H "Content-Type: application/json" \ -d '{ "model": "cydonia-24b-v4.3", "messages": [ {"role": "system", "content": "You are Mara, a lighthouse keeper. Stay in character."}, {"role": "user", "content": "The storm has the crossing closed. Can I stay?"} ], "max_tokens": 300 }'from openai import OpenAI client = OpenAI(base_url="https://api.unrestrictable.ai/v1", api_key="sk-unr-live-…")reply = client.chat.completions.create( model="cydonia-24b-v4.3", messages=[{"role": "user", "content": "Describe the tavern as the door opens."}], max_tokens=300,)print(reply.choices[0].message.content)For SillyTavern, choose Chat Completion with a custom endpoint and paste an id from the table; the field names are in the SillyTavern section of the docs. Janitor AI and the other front-ends on the docs page take the same URL and key.
Running the weights yourself
Each of the five repos is public safetensors in BF16. At that precision a 24B weighs roughly 48 GB, so for one GPU the GGUF linked from the card is the realistic choice, with a quant that fits the card.
# the safetensors repo, then a GGUF for llama.cpp or KoboldCppgit lfs installgit clone https://huggingface.co/TheDrummer/Cydonia-24B-v4.3# the card points to a GGUF repo and recommends bartowski's iMatrix quantsLoad the GGUF in llama.cpp, KoboldCpp, or Ollama via a Modelfile. If you have no GPU, the KoboldAI Lite guide shows the same front-end pointed at this API instead.
Questions
TheDrummer is a Hugging Face account publishing roleplay and creative-writing fine-tunes, among them Cydonia, Magidonia, Skyfall, Anubis, Behemoth and Rocinante. On 27 October 2026 the profile listed 222 models. It publishes weights; it is not itself a chat app.
Yes, five of them: Cydonia 24B v4.3 and v4.1, Magidonia 24B v4.3, Skyfall 36B v2 and Anubis 70B v1.1. Each runs in /chat with three free turns a day and by API, billed per token. Also served here: Anubis 70B v1, Artemis v1.1, Cydonia 24B v2, Rocinante 12B v1.1, UnslopNemo 12B v4.1. See Cydonia 24B v4.3.
According to its card, Cydonia 24B v4.3 is tuned from Mistral Small 3.2 24B Instruct 2506 and uses the Mistral v7 Tekken template. Magidonia 24B v4.3 applies the same tune to Magistral Small 2509.
Rivermind-12B-v1 is a 12B model that appeared on the TheDrummer profile on 27 October 2026. It is not served here; download it from the profile.
Community leaderboards live on Hugging Face, and uncensored builds get scored on the UGI board. Here, a monthly board covers of its own builds with the daily check beside every score. See the uncensored LLM leaderboard.
It compares the token count of a fixed test prompt from one day to the next, which exposes a swapped model or an injected prompt. Two tunes of the same base share a tokenizer, so it cannot separate Cydonia v4.3 from v4.1. Today’s finding is in the verdict column. See how the daily check works.
Every model above is on the uncensored model list with its live price per million tokens. Per-token pricing sets the whole catalog side by side, and the family directory groups the builds by base model.