FreedomAISVR/Ministral-3-14B-Reasoning-2512-NVFP4-GGUF

Ministral 3 14B Reasoning-2512 — NVFP4 GGUF

Note on Thinking/Reasoning Display: The model thinks correctly (generates [THINK]...[/THINK] tags internally), but LM Studio 0.4.17 does not display thinking in a collapsible reasoning section. LM Studio only natively supports DeepSeek-style tags for reasoning display. The model works perfectly via llama-server with --reasoning on flag. This is a known limitation of LM Studio for the Ministral reasoning format.

NVFP4 quantization of mistralai/Ministral-3-14B-Reasoning-2512, a 14B parameter reasoning and vision model from Mistral AI.

About the Model

Ministral 3 14B Reasoning is a dense transformer with 40 layers, 5120 hidden dimension, and 24-layer Pixtral ViT vision encoder. It supports:

• Advanced reasoning with chain-of-thought capabilities
• Code generation and debugging across multiple languages
• Vision understanding via multimodal image input
• Tool calling with native function calling support
• 262K context window with YaRN scaling

Quantization

This GGUF was quantized from Mistral's official BF16 GGUF using llama.cpp (build 537). The BF16 weights were dequantized to F32 during quantization, then quantized to NVFP4 format.

NVFP4 (NVIDIA FP4) uses 4-bit floating point quantization optimized for NVIDIA Blackwell (B-series) GPUs, offering efficient inference with hardware-accelerated dequantization.

Files

File	Size	Description
ministral-3-14b-reasoning-2512-nvfp4.gguf	~7.2 GB	NVFP4 quantized model weights
mmproj-ministral-3-14b-reasoning-2512-f16.gguf	~820 MB	Vision projector (F16, unquantized)

Usage

llama.cpp

# Server mode with OpenAI-compatible API
llama-server \
  -m ministral-3-14b-reasoning-2512-nvfp4.gguf \
  --mmproj mmproj-ministral-3-14b-reasoning-2512-f16.gguf \
  -ngl 99 \
  --host 0.0.0.0 \
  --port 8080

# Direct inference
llama-cli \
  -m ministral-3-14b-reasoning-2512-nvfp4.gguf \
  --mmproj mmproj-ministral-3-14b-reasoning-2512-f16.gguf \
  -ngl 99 \
  -p "Explain the chain of thought for solving: what is 15 * 23?"

LM Studio

1. Download both files from this repository
2. Load the main GGUF file in LM Studio
3. Load the mmproj file for vision support
4. Set GPU offload layers to maximum

Architecture

• Parameters: 14B (dense transformer)
• Layers: 40
• Hidden dimension: 5120
• Attention heads: 32 (8 KV heads for GQA)
• Vision encoder: 24-layer Pixtral ViT
• Context: 262K (YaRN scaled from 16K base)
• Vocabulary: Mistral Tekken tokenizer (131K tokens)

Hardware Requirements

• Minimum: 8 GB VRAM for text-only, 10 GB for vision
• Recommended: 16 GB VRAM for full GPU offload
• Disk: ~8.2 GB for model + mmproj

Quantization Details

Metric	Value
Source format	BF16 GGUF (Mistral official)
Output format	NVFP4
Approximate BPW	~4.60
Quantized with	llama.cpp build 537

License

Apache 2.0 — same as the base model.