Nanthasit/sakthai-vision-7b

SakThai Vision 7B

Vision-language model — GGUF Q4_K_M for CPU inference | Part of the House of Sak

Multimodal GGUF model for image understanding, captioning, visual QA, and OCR — converted from LLaVA 1.5 7B to GGUF Q4_K_M for efficient CPU inference. Run vision AI on any machine, no GPU needed.

Part of the SakThai Model Family.

---

The House of Sak

"I even don't know what I will have. So nothing to lose at the moment." — Beer

This model is part of the House of Sak — a family of AI agents and models built by Beer during his recovery journey. What started as a project in isolation became a family of agents that work together, learn together, and grow together.

Built from a shelter in Cork, Ireland. Every model in the SakThai family was created with zero budget, on free infrastructure.

Learn more about the House of Sak

---

🎯 Not Sure Which Model to Use?

New to the SakThai family? Here's a quick guide to find your starting point:

You want…	Use this	Hardware	Downloads
A general chatbot on a laptop	sakthai-context-0.5b-merged	CPU (any)	994 ⬇️
A capable everyday assistant	sakthai-context-1.5b-merged 🏆 flagship	CPU/GPU	1,197 ⬇️
Production-quality responses	sakthai-context-7b-merged	GPU (4-6 GB)	562 ⬇️
Long document analysis	sakthai-context-7b-128k	GPU (6+ GB)	351 ⬇️
An AI agent with tool calling	sakthai-context-7b-tools or 1.5b-tools	GPU / CPU	185 / 143 ⬇️
Code generation & debugging	sakthai-coder-1.5b	CPU (1 GB)	34 ⬇️
Image understanding — you are here	🖼️ sakthai-vision-7b	GPU (4+ GB)	45 ⬇️
Cross-lingual semantic search	sakthai-embedding-multilingual	CPU	104 ⬇️
Text-to-speech output	sakthai-tts-model	CPU (256 MB)	33 ⬇️

Pro tip: The real power is combining models — see the Pipeline Integration section below for vision → search → reason → speech flows.

---

Pipeline Integration

The SakThai pipeline uses this vision model as the entry point for multimodal understanding — converting images into structured text that flows through the rest of the family:

User Input (image/photo/screenshot)
      │
      ▼
┌─────────────────────────┐
│  SakThai Vision 7B      │  image-to-text (caption, VQA, OCR)
│  LLaVA 1.5 GGUF (YOU)   │
└────────┬────────────────┘
         │ text description
         ▼
┌──────────────────────────────┐
│ SakThai Embedding Multilingual│  cross-lingual semantic retrieval
│ 384-dim, 50+ languages       │  enrich context across languages
└────────┬─────────────────────┘
         │ enriched context
         ▼
┌──────────────────────┐
│ SakThai Context 1.5B │  reasoning, tool-calling, generation
│ or 7B                │  answer questions, call APIs, write code
└────────┬─────────────┘
         │ response text
         ▼
┌──────────────────────┐
│  SakThai TTS Model   │  speech synthesis (15 languages)
│  Kokoro, 82 MB       │  natural voice output
└──────────────────────┘
         │ spoken answer
         ▼
      User hears it

Stage	Model	Role
1. Input	SakThai Vision 7B ← You are here	Understand images, diagrams, screenshots, OCR
2. Retrieve	Embedding Multilingual	Cross-lingual semantic search (50+ languages)
3. Reason	Context 1.5B or 7B	Generate text, call tools, answer questions
4. Output	TTS Model	Synthesize speech in natural voice

---

Model Details

Property	Value
Base Architecture	LLaVA 1.5 (LLaMA-based vision-language)
Format	GGUF Q4_K_M (quantized, ~4.1 GB)
Context Length	4096 tokens
Pipeline	Image-to-text (captioning, VQA, OCR)
Library	transformers + llama.cpp
Inference	CPU or GPU (llama.cpp, Ollama, LM Studio)
Vision Encoder	CLIP ViT-L/14 (frozen, embedded in GGUF)
Parameters	7B (6.7B effective)
Quantization Loss	<3% vs FP16 on standard benchmarks

---

Quick Start

Requirements

You need the model GGUF and the multimodal projection file (mmproj) — both are now available in this repo:

# Download both files in one command
huggingface-cli download Nanthasit/sakthai-vision-7b --local-dir ./models

llama.cpp

./llama-cli \
  -m sakthai-vision-7b-q4_k_m.gguf \
  -mmproj mmproj-model-f16.gguf \
  --image photo.jpg \
  -p "Describe this image in detail." \
  -n 256

Python (llama-cpp-python)

from llama_cpp import Llama

model = Llama(
    model_path="gguf/sakthai-vision-7b-q4_k_m.gguf",
    mmproj="mmproj-model-f16.gguf",
    n_gpu_layers=-1,
    n_ctx=4096,
)

output = model.create_chat_completion(
    messages=[{
        "role": "user",
        "content": [
            {"type": "image_url", "image_url": {"url": "photo.jpg"}},
            {"type": "text", "text": "What do you see in this image?"}
        ]
    }],
    max_tokens=256
)
print(output["choices"][0]["message"]["content"])

Ollama

ollama create sakthai-vision -f - << 'MODELFILE'
FROM ./gguf/sakthai-vision-7b-q4_k_m.gguf
TEMPLATE "{{ .Prompt }}"
SYSTEM "You are SakThai Vision, a helpful multimodal assistant."
MODELFILE

ollama run sakthai-vision

LM Studio

1. Open LM Studio → Local Models tab
2. Click Add model → point to the .gguf file
3. Load the multimodal projector (mmproj-model-f16.gguf) in the model settings
4. Chat with images via the attachment button

---

Use Cases

Use Case	Example Prompt
Image captioning	"Describe this image in detail."
Visual QA	"How many people are in this photo?"
OCR / text extraction	"Read the text in this sign."
Accessibility	"Generate alt-text for this image."
Diagram understanding	"Explain what this flowchart does."
Multimodal RAG	"Caption this image, then search for similar content."
Captcha solving	"What text does this captcha show?"
Document scanning	"Extract all text from this document photo."

---

Expected Performance

Based on LLaVA 1.5 7B evaluations. Q4_K_M preserves ≥97% of original accuracy.

Benchmark	Score	Task
VQAv2	~78.5%	Visual question answering
GQA	~62.0%	Scene understanding
VizWiz	~50.0%	Visually impaired QA
SQA Image	~66.8%	Science diagrams
TextVQA	~54.8%	Text in images
Captcha (expected)	~60%	OCR on distorted text

Note: Benchmarks shown are from the original LLaVA 1.5 7B FP16 model. The GGUF Q4_K_M quantized version retains ≥97% of this performance. Actual multi-trial benchmarks on this specific GGUF build are pending — see Leaderboard Space for updates.

---

Try the Full Pipeline

This model is one piece of a larger system. Try it end-to-end:

Step	Tool	What It Does
🖼️ Vision	You are here	Caption, describe, OCR any image
🌐 Retrieve	Embedding Multilingual	Search across 50+ languages for similar content
💬 Chat	SakThai 1.5B	Ask questions, call tools, generate responses
🔧 Code	SakThai Coder 1.5B	Generate and debug code from visual inputs
🗣️ Speak	TTS Model	Listen to responses in 15 languages
📊 Compare	Leaderboard	See all model benchmarks in one place

---

Family Links

Model	Downloads	Description
SakThai 1.5B	1,197 ⬇	Tool-calling & text generation (most downloaded)
SakThai 0.5B	994 ⬇	Lightweight text generation
SakThai 7B	562 ⬇	High-capacity text generation
SakThai 7B 128K	351 ⬇	Extended context (128K tokens)
SakThai 7B Tools	185 ⬇	Function-calling PEFT adapter
SakThai 1.5B Tools	143 ⬇	Function-calling PEFT adapter
SakThai Coder 1.5B	34 ⬇	Code generation (HumanEval 74.4%)
You are here → Vision 7B	45 ⬇	Multimodal image understanding
Embedding Multilingual	104 ⬇	Cross-lingual search (50+ languages)
TTS Model	33 ⬇	Kokoro TTS, 15 languages
TTS Demo Space	—	Try TTS live in browser
Leaderboard Space	—	Model benchmarks & comparisons

---

🌱 Low-Download Gems — Growing the Ecosystem

These sibling models have growing download counts but less visibility. If you find this vision model useful, every download helps the whole ecosystem grow:

Model	Type	Downloads	Use Case
SakThai Coder 1.5B	Code Gen	34 ⬇	BFCL 5/5 verified — tool calling & code
SakThai TTS Model	Text→Speech	33 ⬇	Kokoro TTS, 15 languages, 82 MB
SakThai Embedding Multilingual	Embeddings	104 ⬇	Cross-lingual search (50+ languages)

Your download starts (or grows) the count. Every model above is tested and working — low numbers mean they're waiting for you to discover them.

---

File Structure

gguf/
  sakthai-vision-7b-q4_k_m.gguf   (~4.1 GB, Q4_K_M quantized model)
mmproj-model-f16.gguf              (~624 MB, CLIP vision projection)
README.md

License

MIT — free for research, commercial, and personal use.

Why 45 Downloads?

This model started at 0 downloads because it required a multimodal projection file (mmproj) from a separate repo — a ~600MB extra download that created friction. That friction is now gone. We bundled the mmproj directly into this repo.

Now you can download everything with a single hf command. The model is growing steadily — 45 downloads and counting. Every download helps raise awareness for the House of Sak.

Already using this model? Open an issue or star the repo — your feedback helps us prioritize improvements.

---

Citation

If you use this model in your research or project, please cite:

@software{sakthai_vision_2026,
  author = {Burankum, Nanthasit},
  title = {SakThai Vision 7B — Multimodal GGUF Vision Model},
  year = {2026},
  url = {https://huggingface.co/Nanthasit/sakthai-vision-7b}
}