TheDrummer: Skyfall 36B V2

TheDrummer: Skyfall 36B V2 is a text generation model from TheDrummer, released on 10 March 2025. It is an enhanced iteration of Mistral Small 2501, fine-tuned specifically for improved creativity, nuanced writing, role-playing, and coherent storytelling. The model's knowledge cutoff is 30 June 2024, so it works from information available up to that point.

The model accepts text input and returns text output, with a context window of 32,768 tokens and a maximum output of 29,491 tokens. It supports a broad set of sampling and control parameters: temperature, top_p, top_k, frequency_penalty, presence_penalty, repetition_penalty, max_tokens, stop, seed, logit_bias, logprobs, top_logprobs, response_format, and structured_outputs. That combination gives you both creative latitude through the sampling parameters and tighter control through structured outputs and response formats when you need predictable shapes.

The fine-tuning focus points to a few practical uses. Creative and nuanced writing tasks fit well, such as drafting fiction, expanding outlines into prose, or producing varied marketing copy. Role-playing and character-driven dialogue are also a natural match, whether for interactive fiction, game NPCs, or chat personas that need to stay in character over a long exchange. Coherent storytelling across longer passages is another use, helped by the 32,768 token context, which lets you keep prior scenes, character notes, or style guides in the prompt. Structured outputs and response_format make it usable for tasks where the reply must follow a schema, such as extracting fields from a narrative or generating consistent JSON for downstream code.

You call the model through api14's OpenAI-compatible API using the model id "thedrummer/skyfall-36b-v2". Because the interface follows the OpenAI chat completions shape, you can point an existing OpenAI client at api14's base URL, keep your usual request structure, and set the model field to that id. The request parameters listed above are supported, so you can pass temperature, top_p, top_k, frequency_penalty, presence_penalty, repetition_penalty, max_tokens, stop, seed, logit_bias, logprobs, top_logprobs, response_format, and structured_outputs as needed. The live price table on this page covers cost.

Two limits are worth noting when choosing. Input is text only, so if your task depends on images or audio, this model is not the fit. The 32,768 token context also caps how much prior material you can keep in a single request, which matters for very long documents or extended conversations; in those cases you would need chunking or retrieval, or a model with a larger context window.