Devstral-Small-2-NVFP4-GGUF

GGUF quantization of mistralai/Devstral-Small-2-24B-Instruct-2512 — a 24B dense multimodal coding model built on Mistral-Small-3.1 with 393K context, vision support, and tool calling.

Quantized to NVFP4 format for efficient inference with minimal quality loss.

About NVFP4

NVFP4 is NVIDIA's native 4-bit floating-point format (E4M3) for Blackwell GPUs. It stores weights in FP4 with a shared per-block scale, enabling native Blackwell tensor core acceleration with no dequantization overhead during inference.

Files

Quantization Details

Model Description

• Developer: Mistral AI
• Base Model: Mistral-Small-3.1-24B-Base-2503
• Architecture: Dense transformer, Mistral3ForConditionalGeneration
• Parameters: 24B
• Context Length: 393,216 tokens (native)
• Vision: 24-layer SigLIP ViT (1024 hidden), image-text-to-text
• Capabilities: Coding, tool calling, agentic workflows
• Languages: Multilingual
• License: Apache 2.0

Usage

llama.cpp (CLI)

# Text + Image
llama-cli -m devstral-small-2-nvfp4.gguf \
  --mmproj mmproj-devstral-small-2-f16.gguf \
  --image photo.jpg \
  -p "Describe this image in detail" \
  -n 512

# Text only
llama-cli -m devstral-small-2-nvfp4.gguf \
  -p "Write a Python function to sort a list" \
  -n 512

# OpenAI-compatible server
llama-server -m devstral-small-2-nvfp4.gguf \
  --mmproj mmproj-devstral-small-2-f16.gguf \
  --port 8080

llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="FreedomAISVR/Devstral-Small-2-NVFP4-GGUF",
    filename="devstral-small-2-nvfp4.gguf",
    n_gpu_layers=-1,
)

response = llm.create_chat_completion([
    {"role": "user", "content": "Write a Python function to sort a list"}
])
print(response["choices"][0]["message"]["content"])

Direct download

from huggingface_hub import hf_hub_download

for filename in ["devstral-small-2-nvfp4.gguf", "mmproj-devstral-small-2-f16.gguf"]:
    hf_hub_download(
        repo_id="FreedomAISVR/Devstral-Small-2-NVFP4-GGUF",
        filename=filename,
        local_dir="./models"
    )

Quantization Pipeline

1. Download source weights
   huggingface_hub.snapshot_download("mistralai/Devstral-Small-2-24B-Instruct-2512")

2. Convert text model to F16 GGUF (dequantize FP8)
   convert_hf_to_gguf.py --outtype f16

3. Extract vision encoder
   convert_hf_to_gguf.py --mmproj --outtype f16

4. Quantize to NVFP4
   llama-quantize devstral-small-2-f16.gguf devstral-small-2-nvfp4.gguf NVFP4

Hardware

License

Apache 2.0 — same as the original mistralai/Devstral-Small-2-24B-Instruct-2512.