Lumimaid-Nexus gets a classic Mixtral conversion for narrative-model users
The community release repackages an existing narrative-focused MoE merge for engines expecting per-expert tensors, without changing weight values according to its author.
A compatibility release for the narrative merge
Community model creator mrcuddle has published a classic Mixtral-layout conversion of Lumimaid-Nexus-2x12B-MoE. The October 14 repository targets inference engines that expect separate expert tensors, including vLLM, llama.cpp conversion tools and older Transformers, according to its English model card.
That is a practical packaging change for users interested in the existing narrative-model merge. The original combines NeverSleep’s Lumimaid-v0.2-12B with ReadyArt’s Dark-Nexus-12B-v2.0 through mergekit-moe. Its author describes the donors’ focuses as narrative consistency and character interaction. Those descriptions are intended use cases, not independently established quality rankings.
Same values, different tensor layout
The source checkpoint uses the Transformers-v5 packed MoE format, stacking expert tensors within each layer. The new repository unpacks those into classic block_sparse_moe.experts tensors, splitting gate and up projections into separate per-expert entries. The author explicitly says no numerical values were changed, only layout and naming.
The card also says the model configuration, tokenizer and chat template are identical to the source. Users should retain the source’s documented ChatML/Tekken prompt format. The repository contains BF16 Safetensors shards; this release is neither a GGUF quantization nor a newly trained model, and its two-expert name should not be read as a new capability benchmark.
Hugging Face metadata dates the original merge repository to October 8 and the conversion repository to October 14. The conversion and source cards declare Apache-2.0. Running the downloaded weights still requires a compatible inference stack and sufficient hardware; the conversion card does not provide measured memory requirements or a tested engine-version matrix.
AITrending reviewed the English cards, repository metadata and configuration. It has not downloaded the weights, compared their numerical contents or verified serving behavior. The news is the newly available compatibility package, with the author’s claims clearly separated from independent testing.