VRAMGlass contains affiliate links. As an eBay Partner, we may be compensated if you make a purchase.As an eBay Partner, we may be compensated.Learn more

GPUs tracked57Price history49 daysPrice points (24h)43,400 pointsBiggest 24h dropGeForce RTX 3060 12GB-11.1%Biggest 24h riseCMP 170HX 40GB+19.8%Updated
$ USD

How to Deploy Ministral 3 14B Locally

Ministral 3 14B is the largest of the small Mistral 3 models; it understands images, and Mistral says it performs close to Small 3.2 24B. Yet it needs about 6 GB less VRAM: 4-bit with 8K takes only about 10.6 GB, so a 16GB card runs it with room left for more than 40K of context. Context tops out at 256K, and it is Apache 2.0 licensed.

Ministral 3 14B details

  • Min VRAM at 4-bit11 GB · 8K
  • Parameters13.5B
  • Max context256K
Architecture
Dense
License
Apache-2.0
Launch date
Dec 2, 2025
Vendor
Mistral AI
Hugging Face repo
mistralai/Ministral-3-14B-Instruct-2512
GGUF repo
mistralai/Ministral-3-14B-Instruct-2512-GGUF
NVFP4 repo (Blackwell)
RedHatAI/Ministral-3-14B-Instruct-2512-NVFP4
MLX repo (Apple)
mlx-community/Ministral-3-14B-Instruct-2512-4bit
Ollama tag
ministral-3:14b
Revision
2512
Layers
40
Hidden size
5,120
KV heads
8
Head dim
128

How much VRAM does Ministral 3 14B need for local deployment?

QuantizationWeightsGB8K contextGB32K contextGB128K contextGB256K context (max)GB
1-bit3.7GGUF6.110.126.247.7
2-bit5.5GGUF7.911.92849.5
3-bit6.9GGUF9.313.329.450.9
4-bit8.2GGUF10.614.630.752.2
5-bit9.6GGUF121632.153.6
8-bit14.4GGUF16.820.836.958.4
BF16/FP1627GGUF29.433.449.571

Full calculation rules are on the Methodology page.

Can a single GPU run Ministral 3 14B?

Yes, and a 16GB card is enough to make it your everyday model: 4-bit with 32K of context comes to about 14.6 GB. That's also as far as 16GB goes, though, since 8-bit weights alone nearly fill the card and leave almost no room for context. On a 24GB card, 8-bit is the better pick, and Mistral itself says its official weights fit in 24GB as published; an RTX 3090 runs it at about 45 tokens/s with roughly 60K of context. The advertised 256K window needs about 52 GB, which only a 96GB card or a 128GB unified-memory machine can hold in full.

eBay
DevicePriceSpeed t/sMax context
Tesla V100 16GB16 GB$37024 listings70.247K
GeForce RTX 5060 Ti 16GB16 GB$86958 listings36.847K
Tesla V100 32GB32 GB$89518 listings70.2149K
Radeon RX 7900 XTX24 GB$1,25029 listings74.498K
GeForce RTX 309024 GB$1,799107 listings72.898K
GeForce RTX 409024 GB$3,75868 listings77.798K
Ryzen AI Max+ 395 128GBWhole system · 128 GB$4,46420 listings21.5256K
RTX A600048 GB$4,89916 listings60.8251K
Mac Studio M5 Ultra 96GBWhole system · 96 GB$7,27412 listings92.6256K
MacBook Pro M5 Max 128GBWhole system · 128 GB$7,54920 listings49.5256K
DGX Spark 128GBWhole system · 128 GB$8,24022 listings22.9256K
GeForce RTX 509032 GB$8,395123 listings127.4149K
RTX PRO 600096 GB$22,50011 listings127.4256K

Only common GPUs and systems are listed, one set for models under 50B and another for larger ones, sorted by price from low to high.

What to know before running Ministral 3 14B locally

  1. Instruct or Reasoning

    This size comes in three versions, Base, Instruct and Reasoning, and all of them understand images. Instruct answers directly and is released in FP8; Reasoning thinks before answering and is released in BF16, and Mistral says it reaches 85% on the AIME 2025 math competition problems. Use Instruct for everyday chat and tool calling, and switch to Reasoning for hard math and coding problems.

  2. How many bits on a 24GB card

    The GGUFs are about 8.2 GB at 4-bit, 9.6 GB at 5-bit and 14.4 GB at 8-bit. On a 24GB card, we recommend going straight to 8-bit: with 32K it's about 20.8 GB and runs at about 45 tokens/s on an RTX 3090, with context reaching about 61K at most. Mistral also says the original FP8 fits in 24GB of VRAM.

  3. Deployment parameters

    vLLM needs 0.12.0 or later, and tool calling needs --enable-auto-tool-choice and --tool-call-parser mistral. Sampling advice is the same as for the 8B: temperature below 0.1, plus a system prompt that spells out the use case. The Ollama tag is ministral-3:14b.

  4. The catch with long context

    It uses standard attention, so every extra 1K of context adds about 0.17 GB of VRAM. At 128K, the context cache is about 21 GB, more than twice the size of the 4-bit weights; for long-document work, figure out how much context you need first, then pick the quant.

Launch commands

Ollama

  • Windows
  • macOS
  • Linux
Terminal
ollama run ministral-3:14b

llama.cpp

  • NVIDIA
  • AMD
  • Mac
  • CPU
Terminal
llama-server -hf mistralai/Ministral-3-14B-Instruct-2512-GGUF:Q4_K_M \
  -c 8192

vLLM

  • Linux
  • NVIDIA
  • AMD
Original weights
Terminal
vllm serve mistralai/Ministral-3-14B-Instruct-2512
NVFP4 · Blackwell

Blackwell GPUs (RTX 50 series, RTX PRO 6000, B200) run NVFP4 weights natively, at about 30% of the original size.

Terminal
vllm serve RedHatAI/Ministral-3-14B-Instruct-2512-NVFP4

MLX

  • Apple Silicon

Vision models run on mlx-vlm: on a Mac, pip install mlx-vlm first; this command downloads the MLX weights from Hugging Face and starts a server.

Terminal
mlx_vlm.server --model mlx-community/Ministral-3-14B-Instruct-2512-4bit

File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology

Other models in the Mistral family

ModelTotal paramsActive params4-bit size
Mistral 7B v0.3Dense7.3B7.3B4.4 GB
Ministral 3 8BDense8.4B8.4B5.2 GB
Mistral NemoDense12.3B12.3B7.5 GB
Mistral Small 3.2 24BDense24B24B14.3 GB

About Mistral