MythoMax 13B

MythoMax 13B is a text model from Gryphe, identified on api14 as gryphe/mythomax-l2-13b. It is a fine-tune of Llama 2 13B, produced by merging rather than by a single training run, and Gryphe describes it as one of the highest performing and most popular fine-tunes of that base model. It sits in the text category, so it is built for language work rather than for images, audio or other modalities.

The model takes text as input and returns text as output. It has a context window of 8192 tokens, which covers a reasonably long prompt plus the generated reply, and its maximum output is 3686 tokens. Its knowledge cutoff is 30 June 2023, so anything after that date is outside what it learned during training. Gryphe highlights rich descriptions and roleplay as particular strengths of this fine-tune, which is a useful signal about the kind of prose it tends to produce.

Those characteristics point to a fairly specific set of uses. The roleplay and rich description strengths make it a natural fit for interactive fiction, character driven chat, and narrative or dialogue generation where tone and detail matter. The 8192 token context lets you keep a long running conversation or a substantial scene in view without trimming constantly, and the 3686 token output ceiling supports replies that run to several paragraphs. Beyond creative writing, it works for general text generation and chat style assistants, though the knowledge cutoff means you should not rely on it for questions about events after June 2023.

You call it through api14's OpenAI-compatible API using the model id gryphe/mythomax-l2-13b, so existing OpenAI client libraries and request shapes work with only the model name changed. The supported request parameters are frequency_penalty, logit_bias, logprobs, max_tokens, min_p, presence_penalty, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, top_a, top_k, top_logprobs and top_p. That is a wide sampling and control surface: the usual temperature, top_p, top_k and penalty settings, plus min_p, top_a and repetition_penalty, which are common in community fine-tune workflows. Seed gives you reproducible sampling, stop sequences let you cut generation at a chosen string, and logprobs with top_logprobs expose token level probabilities. response_format and structured_outputs are available if you need the output shaped in a particular way.

The main limits to keep in mind are the 8192 token context and the 3686 token output cap, along with the June 2023 knowledge cutoff. If your task needs a much larger context, more recent knowledge, or non text input, a different model on api14 will fit better. For creative writing, character chat and general text generation within those bounds, this model is a reasonable choice.