BigBlueCeiling/FrndoBrain-1.0.1-24b
BigBlueCeiling/FrndoBrain-1.0.1-24b is a 24-billion-parameter multimodal (text + image) instruction-tuned language model, a LoRA fine-tune of Mistral Small 3.2 24B Instruct. Developed by BigBlueCeiling, it maintains the original architecture and vision capabilities while adapting language-model weights towards a specific domain. This model is designed as a drop-in replacement for the base Mistral Small 3.2 24B, offering refined conversational style on in-domain prompts.
Loading preview...
Overview
BigBlueCeiling/FrndoBrain-1.0.1-24b is a 24-billion-parameter multimodal (text + image) instruction-tuned language model, derived from mistralai/Mistral-Small-3.2-24B-Instruct-2506. It is a LoRA fine-tune where only the language-model attention and MLP weights were adapted, then merged back into a full BF16 HuggingFace checkpoint. The model retains the original vision tower, multimodal projector, tokenizer, and overall architecture, making it a byte-identical structural twin to its base.
Key Capabilities & Features
- Drop-in Replacement: Functions identically to the vanilla Mistral Small 3.2 24B Instruct, requiring no changes to serving commands, prompt templates, or inference pipelines beyond the model path.
- Preserved Multimodality: Image input and vision behavior are fully preserved, as the Pixtral vision encoder and multimodal projector were frozen during training.
- Domain-Specific Adaptation: Language model behavior is shifted towards the style and conventions of the fine-tuning dataset, offering refined responses on in-domain prompts while remaining close to the base model on out-of-domain queries.
- Efficient Fine-tuning: Achieved through QLoRA with a small percentage of trainable parameters (0.38%), ensuring a bounded deviation from the base model's core capabilities.
Ideal Use Cases
- Users already deploying Mistral Small 3.2 24B who seek a refined conversational style for specific domains without altering their existing infrastructure.
- Applications requiring multimodal capabilities (text + image) with a preference for a particular conversational nuance or style.
- Scenarios where a memory-efficient fine-tune is preferred, as the LoRA merge process was optimized for memory constraints.