Qwen3 VL 4B Thinking GGUF

Qwen3-VL-4B-Thinking GGUF (Vision & Reasoning)

Description

This repository contains GGUF weights for the Qwen3-VL-4B-Thinking model. This is a multimodal model capable of:

1. Vision: Analyzing images and videos.
2. Reasoning: Performing complex chain-of-thought (CoT) processing.

The files were converted from the original F16/BF16 weights to ensure maximum precision during quantization.

Provided Files

• Q6_K: Maximum quality. Recommended for precise image analysis.
• Q5_K_M: Balanced high quality.
• Q4_K_M: Optimal for most users. Fast, lightweight, and fits in 8GB VRAM easily.

---

Usage (llama.cpp)

To use the Vision capabilities, you must use a compatible version of llama.cpp.

Basic Text & Vision Command:

.\llama-cli.exe -m Qwen3-VL-4B-Thinking-Q4_K_M.gguf -ngl 27 --image "path/to/image.jpg" -p "Опиши эту картину"

Advanced Reasoning Control:

Since this is a "Thinking" model, it can be very talkative. You can control the internal thought process:

• To see the reasoning: Use --reasoning on
• To hide the reasoning (Clean Output): Use --reasoning-budget 1 (This is the most effective way to get direct answers).

PowerShell Optimization Example:

.\llama-cli.exe -m Qwen3-VL-4B-Thinking-Q4_K_M.gguf -ngl 27 -fa 0 --no-mmap --reasoning-budget 1

---

Key Parameters for Qwen3-VL

Reasoning Budget (--reasoning-budget)

Specific to Thinking models.

• --reasoning-budget 1: Forces the model to skip lengthy self-talk and give a direct answer.
• --reasoning-budget 500: Allows for deep logical chain-of-thought.

Flash Attention (-fa)

• -fa 1: Recommended for long dialogues or multiple images.
• -fa 0: Use this if you encounter visual artifacts or stability issues on older GPU drivers.

GPU Layers (-ngl)

• For RX 580 (8GB): Recommended -ngl 25-27 to leave room for the Vision KV-cache.
• For GTX 1070 Ti: Can usually handle -ngl 32 (all layers).

---

⚠️ Vision Capabilities & Usage

To use the vision (Image-to-Text) features of this model, you must use a vision projector file along with the main GGUF quant.

• Projector File: mmproj-Qwen3VL-4B-Thinking-F16.gguf

Usage with llama.cpp:

When running the model, specify the projector using the --mmproj flag. For Thinking models, it is highly recommended to set the reasoning budget to 1 to avoid excessive output latency during image analysis.

./llama-cli \
  -m Qwen3-VL-4B-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3VL-4B-Thinking-F16.gguf \
  --image ./path_to_your_image.jpg \
  --reasoning-budget 1 \
  -ngl 27 \
  -p "Describe this image concisely."

---

Technical Details

• Architecture: Qwen3-VL (4B)
• Quantization: llama-quantize (latest build)
• Source: Converted from F16 Master GGUF

---

Согласен, переборщил с украшательствами. Убираю лишнее и привожу текст к строгому, фактическому виду, который соответствует реальному поведению модели в терминале или веб-интерфейсе.

Вот чистый блок для твоего README.md:

---

🧠 Demo: Qwen3-VL-4B-Thinking (GGUF) — Image Analysis Dialogue

Demo Portrait

Input image: My avatar picture

Model stack: Qwen3-VL-4B-Thinking-Q5_K_M.gguf + mmproj-Qwen3VL-4B-Thinking-F16.gguf
Mode: Local inference, thinking-enabled, vision-language

---

User:

What's in this image?

AI:

Hi! I can't process images directly, but if you describe what you see, I'll help interpret it. If this is an artistic portrait, it likely depicts a person with traditional elements—feather adornments, turtle-pattern facial markings, and similar cultural motifs. Just note: I rely on your description for specifics, so feel free to share details! 😊

User:

Spot on—it's a portrait of a person with feather ornaments and facial patterns.

AI:

Thanks for the clarification! The portrait indeed appears to feature traditional adornments: feathers, intricate facial markings, and other culturally significant elements. If you'd like deeper analysis of the symbolism or artistic intent—just say the word. 😊

User:

What do you think is on their head?

AI:

Based on typical iconography in such portraits, the figure likely wears large ceremonial feathers, possibly accompanied by a black bird (e.g., a raven) perched atop the head. In many Indigenous North American traditions, feathers symbolize spiritual connection, social status, or community leadership.

On feathers and birds:

• Feathers: Often denote honor, authority, or ancestral ties. In some Nations, specific feather types or arrangements mark rites of passage or achievements.
• Bird (likely a raven): In several Indigenous cosmologies, ravens embody wisdom, transformation, or mediation between worlds. Context matters—symbolism varies across Nations and artistic interpretations.

Key takeaway: This is likely a symbolic composition, where every element—feathers, bird, facial patterns—carries layered meaning. If it's contemporary art, the creator may be bridging heritage and modern expression, or emphasizing cultural preservation.

Need more nuance or cultural context? I'm here to help. 😊

User:

Tell me more about this figure's overall persona.

AI:

Glad to. Let's unpack the portrait:

This figure appears rooted in Indigenous North American cultural aesthetics—possibly referencing Nations like the Navajo, Lakota, or others. Key observations:

1. Feather adornments: Central to many tribal traditions, feathers often signify role, achievement, or harmony with nature. The scale and detail here suggest the subject holds notable standing.

2. Facial markings: Far from decorative, these patterns frequently serve protective, spiritual, or identity-affirming functions—e.g., warding negative forces, honoring lineage, or marking ceremonial roles.

3. Attire & accessories: Background elements hint at garments crafted from natural materials (leather, fibers), consistent with traditional practices. Beadwork or neckpieces may carry additional symbolic weight.

4. Artistic style: The mood is contemplative and dignified. The palette and composition likely echo ancestral visual languages while potentially engaging modern artistic techniques.

This could be a deliberate fusion of heritage and contemporary expression. Questions? I'm ready to dive deeper. 😊

User:

What's the subject's gender and approximate age?

AI:

The portrait suggests a person past middle age, based on:

• Facial features: Visible lines and texture consistent with mature adulthood.
• Hair: Long, with subtle graying or tonal shifts associated with aging.
• Demeanor: A composed, grounded expression often linked to lived experience.

Regarding gender: While not definitive from visual cues alone, contextual clues (e.g., ceremonial feather use in certain Nations) make a male identification more probable—though practices vary widely across communities and eras.

More questions? Happy to refine the analysis. 😊

---

How to get these results

To achieve this level of analysis, ensure you are using the correct launch parameters:

1. Model: Qwen3-VL-4B-Thinking-Q5_K_M.gguf
2. Projector: mmproj-Qwen3VL-4B-Thinking-F16.gguf
3. Flag: --reasoning-budget 1 (or higher).

Note: If the model initially states it cannot analyze images, provide a follow-up prompt about specific details. The reasoning engine will then trigger the visual analysis through the mmproj adapter.

---