DAC6.5

DAC6.5 — M.INC. Architecture Series

<p align="center"> <img src="DAC6.5_Logo/DAC6.5_Logo.jpg" alt="DAC6 Logo" width="800px"/> </p>

M.INC. is proud to present DAC6.5, the flagship release introducing our custom, proprietary multimodal architecture. Designed to bring high-performance vision-language capabilities to small-footprint and edge compute environments, DAC6.5 seamlessly unifies perception and reasoning into a single cohesive ecosystem.

DAC6.5 couples the LFM2.5-230M language backbone with the high-resolution SigLIP2 vision encoder through a custom-trained projection module, delivering strong zero-shot image understanding, rapid inference, and minimal memory overhead.

---

Model Architecture & Technical Overview

Unlike fragmented pipelines requiring complex multi-stage orchestration, DAC6.5 is engineered as an integrated, end-to-end multimodal system:

• Language Backbone (LFM2.5-230M): Ultra-lightweight text engine fine-tuned for high throughput and low-latency response generation.
• Vision Encoder (SigLIP2): State-of-the-art visual feature extraction, capturing deep spatial and semantic details from input images.
• Proprietary Projector: A dedicated cross-modal projection layer trained from scratch by M.INC. to map visual representation spaces directly into the language model's embedding space.

---

Repository Structure & Core Artifacts

This repository contains the complete unified weights, configuration schemas, and quantization variants required for deployment:

Weights & Configurations

• model.safetensors — Unified model weights stored with isolated, clean namespaces (language_model.*, vision_encoder.*, and projector.*).
• config.json — Core DAC6.5 architecture specifications and inter-component routing.
• tokenizer.json, tokenizer_config.json, chat_template.jinja — Complete tokenizer configuration and standard chat templates for the LFM engine.
• vision_config.json, processor_config.json — Preprocessing pipelines and parameter definitions for the SigLIP2 vision encoder.

---

☕ Support my research

---

About M.INC.

M.INC. focuses on research and development of custom neural architectures, efficient language models, and accessible multimodal systems. DAC6.5 represents the first milestone in our custom architecture series, establishing a new baseline for compact, locally deployable artificial intelligence.