Skyfall 36B V2 Guide: An Upscaled Mistral Small for Fiction

Skyfall 36B V2 is an odd size, and the oddness is the interesting part. Mistral never released a 36B model. TheDrummer took Mistral Small 2501, a 24B model, made it bigger, and then continued training it for creative writing and roleplay. The model card says so in one line: "an upscaled version of Mistral Small 2501 with continued training for creativity and RP."

That raises the obvious question a developer should ask before spending money on it: does adding parameters to an existing model actually buy anything, and is it worth the larger memory footprint and the shorter context it inherits? This guide works through both.

What an upscale is, and what it is not

An upscale enlarges a trained model rather than training a new one from scratch. The common community approach is to duplicate a span of existing layers so the network gets deeper, then train further so the copied layers stop being redundant and start specialising. The Skyfall card does not describe its exact method, so treat that as the general technique rather than a description of this model.

Two things follow from it. First, the extra capacity is only as good as the continued training that fills it. An upscale with too little training is a slower copy of its base. Second, the upscale does not change the base's architecture-level properties, including its context window and tokenizer. Skyfall is still, underneath, a January 2025 Mistral Small.

Hugging Face lists the weights at 37B parameters in BF16, which is close enough to the name. As with any parameter count, the headline figure says little about quality on its own.

The base: Mistral Small 2501

Mistral released Small 3 on 30 January 2025 as a 24B model under Apache 2.0. Its model card lists a 32K context window and the Tekken tokenizer with a 131K vocabulary. Mistral was explicit that it was trained without reinforcement learning or synthetic data and positioned it as a base for further tuning, which is exactly how TheDrummer used it.

Mistral has since shipped Small 3.1 and 3.2, and 3.2 in particular cut down on runaway repetition. Skyfall predates both. It has been listed on OpenRouter since March 2025, so in model terms it is from an earlier generation.

32K is the real constraint

OpenRouter lists Skyfall 36B V2 with a 32,768-token context, matching its base. For short-form writing that is fine. For long roleplay it is the limit you will hit first.

A long chat with a detailed character card, a lorebook and a running scene history can pass 32K in an evening. Once it does, your front end must either truncate older turns or summarise them, and the model loses details it would otherwise recall. Newer bases in the same family carry 128K-class windows, which is why a sibling like Cydonia 24B V4.1 handles long sessions with less bookkeeping.

If you choose Skyfall anyway, plan the memory strategy up front: a rolling summary, pinned facts, and retrieval for lore. Managing context covers the same techniques in an agent setting, and they transfer directly to long fiction.

Templates: one recommended, one discouraged, one for stories

The card is unusually specific about prompt formats:

• Mistral v7 Tekken — marked highly recommended. Use this by default.
• Metharme — marked not recommended, even though some roleplay front ends default to it.
• Alpaca — described as possibly interesting for choose-your-own-adventure or story-style use.

Check your front end's template setting before judging the model. A mismatched template produces stilted or broken output that has nothing to do with the weights.

Hardware: 36B is an awkward size

The official GGUF repository lists these quantisation sizes:

• Q3_K_M — 17.9 GB
• Q4_K_S — 21.1 GB
• Q4_K_M — 22.4 GB
• Q5_K_M — 26.2 GB
• Q8_0 — 39.2 GB

On a single 24 GB card a Q4 quant technically fits, but it leaves little room for the KV cache, so you either drop to Q3 or accept a shorter working context. A 24B model at Q4 sits comfortably on the same card. That is the practical cost of the upscale: roughly half again as much memory for the same context ceiling. Best model for self-hosting covers how to size this against your hardware.

The repository lists the licence as "other" with no terms on the page. Read the repository files before using it commercially rather than assuming the base's Apache 2.0 carries through.

Skyfall or Cydonia

On OpenRouter at the time of writing, Skyfall lists at $0.55 per million input tokens and $0.80 per million output, against $0.30 and $0.50 for Cydonia 24B V4.1. So Skyfall is the more expensive, older, shorter-context option. It needs a reason to be chosen.

The reason users give is prose. Testimonials on the card are enthusiastic, and some users say they prefer it to the 70B models they had been running. That is subjective, and it is exactly the kind of claim you should test rather than trust:

1. Pick eight scenes that fit comfortably inside 32K.
2. Generate each with Skyfall and with Cydonia at matched sampler settings.
3. Strip the labels, shuffle, and pick a winner for each pair.

If Skyfall wins clearly on short scenes and your sessions stay short, it is a fair choice. If the result is close, or your sessions run long, the newer base wins on context and cost. A/B testing two models covers how to run that comparison without fooling yourself.