MythoMax L2 13B: Why a 2023 Roleplay Merge Is Still Listed

MythoMax L2 13B is one of the oldest models still sitting in most API catalogues, and one of the cheapest. It is not a general assistant, it was never trained for tool use, and its own author has said it should fade away now that newer bases exist. People keep calling it anyway, and the reasons are worth understanding before you decide whether you should.

It was published on Hugging Face by Gryphe in 2023, built on Meta's Llama 2 13B, and aimed squarely at roleplay and storywriting. It is a merge rather than a fine-tune, which explains both why it has a recognisable voice and why it behaves oddly when you ask it to do anything else.

What a merge is, and why this one stuck

A fine-tune trains a base model on new data. A merge trains nothing: it combines the weights of two or more existing fine-tunes of the same base into one model. MythoMax combines two of Gryphe's own Llama 2 models, MythoLogic-L2 and Huginn.

The model card describes an experimental tensor-level method. Each of the model's 363 tensors received its own blend ratio, with more of Huginn mixed into the tensors at the start and end of the network. Gryphe credits that choice with better coherence across the whole model compared with the earlier MythoMix merge it replaced.

The practical outcome was a 13B model with descriptive, character-consistent prose at a size that ran on consumer hardware. OpenRouter's listing still describes it as one of the most popular Llama 2 13B fine-tunes, and character cards written around its voice are one reason it is still called long after the base model aged.

The prompt format matters more than usual

MythoMax expects an Alpaca-style prompt, not a chat template. The card gives this shape:

<System prompt / character card>

### Instruction:
Write Mira's next reply in a chat between Sam and Mira. Write a single reply only.

### Response:

When it is served behind an OpenAI-compatible chat endpoint, the provider converts your messages array into some prompt format. If that conversion does not match what the model learned, you get the classic symptoms: the model writes the user's lines as well as its own, or never stops. The instruction to write a single reply only is in the card for exactly that reason.

If you see runaway replies, check what template your provider applies, set explicit stop sequences such as ### Instruction:, and keep max_tokens modest. Those three changes fix most complaints people attribute to the model itself.

4K context is the real constraint

Llama 2 shipped with a 4,096-token context window, and Fireworks lists exactly that for its MythoMax deployment, shared between input and output. Some catalogues list a larger window for this model; treat that as a provider-side setting and test it before relying on it, because the underlying model was not trained for it.

Four thousand tokens disappears quickly in roleplay. A detailed character card, a system prompt and a few pages of history can fill it before the model writes anything. Front ends handle this by trimming old turns or summarising them, which is the same problem modern agents face at a larger scale. Context length vs effective context explains why even the advertised window overstates what the model uses well.

The licence you inherit

A merge inherits the terms of its base. MythoMax's Hugging Face card marks the licence as "other" without further detail, so the governing terms are effectively Meta's Llama 2 Community License. That licence permits commercial use for most organisations but requires a separate licence from Meta for products with more than 700 million monthly active users at Llama 2's release date.

For a hobby project that is a non-issue. For a commercial product, read the Llama 2 terms directly rather than assuming the merge is unrestricted. Open-weight model licences compared covers how these community licences differ from MIT or Apache 2.0.

Where it still earns its place

Three situations justify MythoMax in 2026:

• Price. OpenRouter lists it at about $0.08 per million input tokens and $0.11 per million output at the time of writing. For long, chatty sessions that is close to free.
• Continuity. If users built characters around its voice, switching models changes the character. That is a product decision, not just a technical one.
• Local hardware. TheBloke's GGUF, GPTQ and AWQ quantisations make it runnable on a single consumer GPU. Where quantised models lose quality covers what you give up.

Where it does not belong: anything that needs instruction following, factual answers, code, or structured output. Fireworks notes it does not support function calling and is not safety-aligned. If you are moderating user-facing content, that last point means you need your own filter in front of it.

What to try instead

Gryphe's own card points at the obvious move: with Llama 3 released, it says, it is time for MythoMax to slowly fade away. Community roleplay fine-tunes of Llama 3 and 3.1 offer longer context and better instruction following at similar sizes. The Llama 3 8B Lunaris guide covers one such model at a smaller size.

The fair way to decide is a side-by-side on your own material. Take ten real character cards and a few hundred turns of anonymised history, run both models with identical sampling settings, and have someone who knows the characters judge blind. If MythoMax still wins on voice, its price makes it easy to keep. If it loses, the 4K window was the reason to leave anyway.