QuixiAI/Devstral-Vision-Small-2507-gguf

NEW - I exported and added mmproj-BF16.gguf to properly support llama.cpp, ollama, and LM Studio.

Devstral-Vision-Small-2507 GGUF

Quantized GGUF versions of cognitivecomputations/Devstral-Vision-Small-2507 - the multimodal coding specialist that combines Devstral's exceptional coding abilities with vision understanding.

Model Description

This is the first vision-enabled version of Devstral, created by transplanting Devstral's language model weights into Mistral-Small-3.2's multimodal architecture. It enables:

• Converting UI screenshots to code
• Debugging visual rendering issues
• Implementing designs from mockups
• Understanding codebases with visual context

Quantization Selection Guide

Quantization	Size	Min RAM	Recommended For	Quality	Notes
Q8_0	23GB	24GB	RTX 3090/4090/A6000 users wanting maximum quality	★★★★★	Near-lossless, best for production use
Q6_K	18GB	20GB	High-end GPUs with focus on quality	★★★★☆	Excellent quality/size balance
Q5_K_M	16GB	18GB	RTX 3080 Ti/4070 Ti users	★★★★☆	Great balance of quality and performance
Q4_K_M	13GB	16GB	Most users - RTX 3060 12GB/3070/4060	★★★☆☆	The sweet spot, minimal quality loss
IQ4_XS	12GB	14GB	Experimental - newer compression method	★★★☆☆	Good alternative to Q4_K_M
Q3_K_M	11GB	12GB	8-12GB GPUs, quality-conscious users	★★☆☆☆	Noticeable quality drop for complex code

Choosing the Right Quantization

For coding with vision tasks, I recommend:

• Production/Professional use: Q8_0 or Q6_K
• General development: Q4_K_M (best balance)
• Limited VRAM: Q5_K_M if you can fit it, otherwise Q4_K_M
• Experimental: Try IQ4_XS for potentially better quality at similar size to Q4_K_M

Avoid Q3_K_M unless you're VRAM-constrained - the quality degradation becomes noticeable for complex coding tasks and visual understanding.

Usage Examples

With llama.cpp

# Download the model
huggingface-cli download cognitivecomputations/Devstral-Vision-Small-2507-GGUF \
  Devstral-Small-Vision-2507-Q4_K_M.gguf \
  --local-dir .
huggingface-cli download cognitivecomputations/Devstral-Vision-Small-2507-GGUF \
  mmproj-BF16.gguf \
  --local-dir .

# Run with llama.cpp
./llama-cli -m Devstral-Small-Vision-2507-Q4_K_M.gguf \
  -p "Analyze this UI and generate React code" \
  --image screenshot.png \
  -c 8192

With LM Studio

1. Download your chosen quantization
2. Load in LM Studio
3. Enable multimodal/vision mode in settings
4. Drag and drop images into the chat

With ollama

# Create Modelfile
cat > Modelfile << EOF
FROM ./Devstral-Small-Vision-2507-Q4_K_M.gguf
PARAMETER temperature 0.7
PARAMETER num_ctx 8192
EOF

# Create and run
ollama create devstral-vision -f Modelfile
ollama run devstral-vision

With koboldcpp

python koboldcpp.py --model Devstral-Small-Vision-2507-Q4_K_M.gguf \
  --contextsize 8192 \
  --gpulayers 999 \
  --multimodal

Performance Tips

1. Context Size: This model supports up to 128k context, but start with 8k-16k for better performance
2. GPU Layers: Offload all layers to GPU if possible (--gpulayers 999 or -ngl 999)
3. Batch Size: Increase batch size for better throughput if you have VRAM headroom
4. Temperature: Use lower temperatures (0.1-0.3) for code generation, higher (0.7-0.9) for creative tasks

Hardware Requirements

Quantization	Single GPU	Partial Offload	CPU Only
Q8_0	24GB VRAM	16GB VRAM + 16GB RAM	32GB RAM
Q6_K	20GB VRAM	12GB VRAM + 16GB RAM	24GB RAM
Q5_K_M	18GB VRAM	12GB VRAM + 12GB RAM	24GB RAM
Q4_K_M	16GB VRAM	8GB VRAM + 12GB RAM	20GB RAM
IQ4_XS	14GB VRAM	8GB VRAM + 12GB RAM	20GB RAM
Q3_K_M	12GB VRAM	6GB VRAM + 12GB RAM	16GB RAM

Model Capabilities

✅ Strengths:

• Exceptional at converting visual designs to code
• Strong debugging abilities with visual context
• Maintains Devstral's 53.6% SWE-Bench performance
• Handles multiple programming languages
• 128k token context window

⚠️ Limitations:

• Not specifically fine-tuned for vision-to-code tasks
• Vision performance bounded by Mistral-Small-3.2's capabilities
• Requires decent hardware for optimal performance
• Quantization impacts both vision and coding quality

License

Apache 2.0 (inherited from base models)

Acknowledgments

• Original model by Eric Hartford at Cognitive Computations
• Built on Mistral AI's Devstral and Mistral-Small models
• Quantized using llama.cpp

Links

• Original Model
• Devstral Base
• Mistral-Small Vision

---

For issues or questions about these quantizations, please open an issue in the repository.