ReMM SLERP L2 13B: Rebuilding MythoMax From Its Parts

ReMM SLERP L2 13B is a model merge from 2023, published on Hugging Face by Undi95. Its model card describes it in one line: a recreation trial of the original MythoMax-L2-13B with updated models. If you already know MythoMax, that tells you most of what to expect. If you do not, the short version is a Llama 2 13B model tuned by merging, aimed at roleplay and fiction rather than tasks.

The interesting part is not the output style, which is close to its parent. It is the recipe, which is unusually well documented, and the licence, which is stricter than people assume.

What ReMM set out to do

Gryphe built MythoMax by merging his MythoLogic-L2 with Huginn, a Llama 2 fine-tune from The-Face-Of-Goonery. By mid-2023 several of the models that had gone into MythoLogic had newer versions. ReMM is an attempt to rebuild the same structure with those newer parts: re-create a MythoLogic-like model from current ingredients, then merge it with Huginn the way MythoMax did.

That makes it a controlled experiment of sorts. Same base, same overall shape, refreshed components. Whether the refresh helps is a question of taste and of your material, which is why the comparison against MythoMax itself is the only test that matters. The MythoMax L2 13B guide covers the parent model, its prompt format and its limits.

The recipe, step by step

The card lays the build out in three parts, all on top of TheBloke/Llama-2-13B-fp16 as the base:

1. Part 1 is a TIES merge of the base with Chronos-Beluga-v2-13B (density 0.42) and airoboros-l2-13b-2.1 (density 0.56).
2. Part 2 is a second TIES merge of the base with Nous-Hermes-Llama2-13b (density 0.30) and the Part 1 result (density 0.70). The output is ReML-L2-13B, the stand-in for MythoLogic.
3. Part 3 is a SLERP merge of Huginn-13b-v1.2 with ReML, which produces ReMM.

The card marks the SLERP step as testing, credits the script it adapted, and notes that ReML itself was a private recreation. In other words, you can read how it was made, but you cannot rebuild every intermediate from public weights alone.

TIES and SLERP in plain terms

Both are ways of combining fine-tunes of the same base without training anything.

TIES starts from the difference each fine-tune made to the base, keeps only the largest changes (the density setting controls how many survive), resolves cases where two fine-tunes pushed the same weight in opposite directions, and adds the result back. It is designed for combining several models at once while limiting interference between them.

SLERP, spherical linear interpolation, blends exactly two models along an arc rather than a straight line between their weights. The usual argument is that this preserves the scale of the weights better than plain averaging. It is the natural choice for the final two-way blend, which is why ReMM uses TIES for the multi-model steps and SLERP at the end.

None of this changes what the base can do. A merge of Llama 2 13B fine-tunes is still a Llama 2 13B model, with that generation's reasoning, knowledge and context limits.

The licence is stricter than MythoMax's

ReMM's Hugging Face card lists its licence as cc-by-nc-4.0, a Creative Commons licence that prohibits commercial use. That sits on top of the Llama 2 Community License the base already carries. MythoMax's card, by comparison, lists its licence as "other".

If you are building a product, that difference is the first thing to settle, before quality or price. A hosted endpoint makes a model easy to call; it does not change the terms the author published. Read the card and the Llama 2 terms, and get advice if the answer matters to your business. Open-weight model licences compared covers how these licences differ in practice.

Running it behind a chat endpoint

The card specifies the Alpaca prompt format: an instruction block headed ### Instruction: followed by ### Response:. Behind an OpenAI-compatible endpoint the provider turns your messages array into some prompt; if that does not match Alpaca, expect the model to write both sides of a conversation or run on past the end of its turn.

Practical defaults that fix most of this:

• Add ### Instruction: as a stop sequence.
• Keep max_tokens modest so a runaway reply costs little.
• Say "write a single reply only" in the system prompt.
• Keep the full prompt well inside the window. OpenRouter lists a 6,144-token context for this model at the time of writing, while Llama 2 was trained at 4,096; test anything past that before relying on it.

The catalogue also lists JSON mode and structured outputs for this model. Those are features of the serving stack, not of the weights; a 2023 creative-writing merge was not trained for schema adherence. Validate any structured output, or use a model built for it.

When ReMM is worth the higher price

OpenRouter lists ReMM at $0.35 per million input tokens and $0.65 per million output at the time of writing. The cheapest MythoMax route on the same catalogue lists $0.08 and $0.11. That is roughly four to six times the price for a model of the same size and generation.

So the case for ReMM is narrow. It is worth calling when you have tested it against MythoMax on your own characters and prefer its voice, and when your use is non-commercial. Otherwise the parent does the same job for less, and a Llama 3 era roleplay model such as the one in the Llama 3 8B Lunaris guide gives you a newer base and a slightly longer 8K window.

To decide, run ten character cards through both 13B merges with identical sampling, strip the model names, and have someone who knows the characters pick blind. If ReMM does not win clearly, the price difference settles it.