MarinaraSpaghetti, also known as Mari, publishes language-model merges and settings for creative conversations. NemoMix-Unleashed combines several Mistral Nemo derivatives and includes the merge configuration, prompting format and sampling recommendations.[1][2][3]
Start with NemoMix-Unleashed12B →
X · Spicy_Marinara_ Contributions
Combine existing models transparently
NemoMix-Unleashed is a merge of pretrained models made with mergekit. Its card names the base and component models and supplies a YAML recipe with their weights and densities. That makes the composition inspectable and credits the upstream creators. It is a model-merging contribution rather than a claim to have pretrained all of the underlying weights.[2]
Settings alongside the weights
A separate SillyTavern Settings repository supplies presets and prompts for users configuring conversational models. Mari’s public portfolio also contains earlier NemoMix versions and other merges. Keeping these companion artifacts available lets users inspect the weight file and part of the setup used to explore its behavior.[1][3]
Ideas
The release notes describe difficulties reconciling special-token identifiers across earlier attempts. Mari provides an updated Mistral instruction format, temperature and sampling suggestions, and links to inference-ready quantizations from other publishers. The practical lesson is that the tokenizer, message format and generation settings can change the experience of a merged model.[2]
Tools, writing & talks
Tools
Start here
NemoMix-Unleashed12Bhuggingface.coA merge of Mistral Nemo-based models with an explicit recipe and usage settings.
Start with the merge recipe and instruction format to understand what changed. Running it requires a compatible inference setup; the card links separately published quantized weights.
Model merge creator and publisher[2]- SillyTavern Settingshuggingface.co
Companion presets and prompts for configuring conversational model frontends.