MythoMax L2 13B

Opper describes MythoMax L2 13B as Gryphe’s 13-billion-parameter open-weight model for creative writing and roleplay, built on Meta’s Llama 2 foundation. It says the model combines MythoLogic-L2 and Huginn through BlockMerge\ Gradient tensor merging, assigning each of 363 tensors a unique blending ratio. The page reports that MythoMax L2 13B uses Alpaca prompt formatting and maintains persona traits, speech patterns, and emotional nuance across long roleplay exchanges. Community GGUF, GPTQ, and AWQ quantizations offer deployment options across different hardware constraints.

MythoMax L2 was released on 2023-08-23 by Gryphe under the Llama 2 Community License Agreement and is documented as a fine-tuned model in the Llama 2 family, targeted at roleplaying and creative writing. It is distributed in several GGUF quantizations, including Q4_K_M, Q5_K_M, Q6_K, and Q8_0, plus an Ollama build. The model synthesizes MythoLogic-L2 and Huginn through a tensor-type merge, applying gradient-tuned blending coefficients across 363 tensors spanning input and output layers. Because each tensor receives a distinct proportion from its parent models, the merge can target robust comprehension from MythoLogic-L2 and immersive writing from Huginn while improving dialogic coherency.

Gryphe's MythoMax L2 13B is an open-source 13B-parameter language model built on the Llama 2 architecture, with a 4K context window and roughly 25.9 GB of required VRAM. The model card lists English proficiency, Alpaca-style input formatting, and a Llama-compatible endpoint format, with the weights hosted on Hugging Face as Gryphe/MythoMax-L2-13b. MythoMax is a MythoMix variant produced by merging MythoLogic-L2 and Huginn with an experimental tensor merge, applying unique ratios across 363 tensors to combine comprehension with writing ability. Reported scores include HellaSwag 83.6, ARC 60.9, and MMLU 55.3, though the catalog shows no evaluation methodology and lists GSM8K as an implausible 9, likely a truncated value.

Gryphe’s MythoMax-L2-13b model card describes a merge of MythoLogic-L2 and Huginn using an experimental tensor-merging method, with unique ratios across 363 tensors and gradient-based adjustment. The creator says this construction increased structural coherence and produced a model proficient at roleplaying and storywriting, while the card also notes that the MythoMax line would slowly fade after Llama 3’s release. The model card says MythoMax-L2-13b primarily uses Alpaca formatting and points users to the merge script and templates used to produce the model. It also identifies GGUF, GPTQ, and AWQ quantized versions by TheBloke, while the Hugging Face listing reports 1,491 downloads in the preceding month and supports use with llama.cpp, LM Studio, and Jan.

gryphe/mythomax-l2-13b (https://huggingface.co/Gryphe/MythoMax-L2-13b); nousresearch/nous-hermes-llama2-13b (https://huggingface.co/NousResearch/Nous ...