MINDI 1.5 is a multimodal agentic AI coding model that generates frontend code (HTML/CSS/JS, Next.js, React, Tailwind) from text prompts and UI screenshots.

Architecture

Component	Details
Base LLM	Qwen/Qwen2.5-Coder-7B-Instruct (7.62B params)
Vision Encoder	CLIP ViT-L/14 (frozen, 304M params)
LoRA Adapters	r=64, alpha=128 (161.5M trainable params)
Fusion	VisionLanguageFusion with text_gate (16.8M params)
Total	8.1B params, 182.5M trainable (2.25%)

Training

3-Phase Progressive Training on AMD MI300X 192GB + Modal A100 40GB
Dataset: 1.45M examples, 860M tokens
Final loss: 0.25–0.40 range

Checkpoint Structure

checkpoints/
├── phase3_final/           ← Best checkpoint for inference
│   ├── lora/               ← LoRA adapter weights
│   ├── vision/             ← Vision projection weights
│   └── fusion/             ← Fusion layer weights
├── phase3_all_step2500_final/
├── phase3_all_step2000/
├── phase3_all_step1500/
└── ... (earlier phases)

Usage

python

from src.model.mindi_model import MINDI15
import torch

model = MINDI15(
    model_name="Qwen/Qwen2.5-Coder-7B-Instruct",
    clip_model="openai/clip-vit-large-patch14",
    hidden_size=3584,
    num_visual_tokens=256,
    torch_dtype=torch.bfloat16,
)
model.load("checkpoints/phase3_final")
model.eval()

response = model.generate(
    prompt="Build a Next.js landing page",
    max_new_tokens=2048,
    temperature=0.7,
)

Special Tokens (22 total, 11 pairs)

think, code, file, critique, suggest, search, error, fix, vision, sandbox, context

Built By

Faaz — MINDIGENOUS.AI | Mumbai, India | April–May 2026

Mindigenous/MINDI-1.5-Vision-Coder

MINDI 1.5 Vision-Coder

Model Description