EldritchLabs/MN-Crimson-Cafe-12B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 15, 2026License:cc-by-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

EldritchLabs/MN-Crimson-Cafe-12B is a 12 billion parameter causal language model, merged from Vortex5/Crimson-Constellation-12B and IggyLux/MN-VelvetCafe-RP-12B-V2 using the multi_fusion method with a cosine_sim importance metric. This model is built on the MistralForCausalLM architecture and features a 32768-token context length. It is designed to combine the strengths of its base models, likely focusing on enhanced conversational or roleplay capabilities given its merge components. The model's configuration specifies a float32 dtype for merging and bfloat16 for output, indicating a balance between precision and efficiency.

Loading preview...

MN Crimson Cafe 12B: A Merged 12B Language Model

MN Crimson Cafe 12B is a 12 billion parameter causal language model developed by EldritchLabs. It is constructed using a sophisticated merging technique to combine the capabilities of two distinct base models: Vortex5/Crimson-Constellation-12B and IggyLux/MN-VelvetCafe-RP-12B-V2.

Merge Details

This model was created using the multi_fusion merge method, which is designed to intelligently combine the strengths of multiple models. The merging process specifically utilized the cosine_sim importance metric, as detailed in a relevant research paper, to guide the fusion of parameters. This approach aims to produce a model that inherits the best characteristics from its constituents.

Technical Configuration

  • Architecture: MistralForCausalLM
  • Base Models: Vortex5/Crimson-Constellation-12B and IggyLux/MN-VelvetCafe-RP-12B-V2
  • Merge Method: multi_fusion with cosine_sim importance metric
  • Parameter Tuning: Includes a tukey_fence of 1.5 during the merge process.
  • Data Types: Merging was performed in float32, with the output model configured for bfloat16 for efficient inference.
  • Tokenizer: Utilizes a union-sourced tokenizer, ensuring broad compatibility with the merged models' tokenization schemes.
  • Context Length: Features a substantial context window of 32768 tokens.

Potential Use Cases

Given its merged nature from models that may have different specializations, MN Crimson Cafe 12B is likely suitable for applications requiring a blend of general language understanding and potentially specific conversational or creative text generation, depending on the original strengths of Crimson-Constellation-12B and MN-VelvetCafe-RP-12B-V2.

Finetunes
1 models