The inputs are WarmBloodAban’s Minimax-h3_Singularity_ref2va_v2_bf16.safetensors and TenStrip’s 10Eros_Max_h3_8stepTURBO_beta6_bf16.safetensors.
The merged weight is distributed as Minimax-h3_Singularity_x_10Eros-Max_PRISM_fp16.safetensors in fp16. Its model card describes a custom PRISM process: randomized-SVD spectral decomposition, polar-geodesic interpolation, sign-conflict mediation, and magnitude re-anchoring. The reported turbo-model share varies by block, from 0.45 to 0.75, with a mean of about 0.55.
The repository lists the license as “other” and says users must check the source-model licenses. It recorded 2,428 downloads in the preceding month and was not deployed by a Hugging Face Inference Provider when the page was captured. The source does not establish output quality, benchmark performance, safety characteristics, or commercial-use rights for the merged model. Source: https://huggingface.co/OliviaRossi/Minimax-h3_Singularity-x-10Eros-Max_PRISM