Ministral-3 14B Instruct 2512

Deploy Ministral-3 14B Instruct 2512 on a dedicated GPU from $0.20 per hour, billed per minute, on 1× RTX A4000. A roughly 14B multimodal instruction model built for practical local and private deployment, with 256K context, native function calling and an FP8 checkpoint that fits in 24 GB VRAM.

Multimodal13.9B paramsApache 2.0Deploys in ~5 minSecure Cloud

mistralai/Ministral-3-14B-Instruct-2512

HexGrid Cloud price

$0.20/hr

1× RTX A4000

Context

256K

tokens

Parameters

13.9B

LM + vision

Precision

FP8

Instruct checkpoint

Running Ministral-3 14B Instruct 2512 on HexGrid Cloud

What you get when you deploy with us, beyond the hourly rate.

Single tenant by default

The model runs on a GPU that is yours for the life of the instance. No shared endpoint, no queue behind other tenants.

Per-minute billing

You pay for the GPU, not the token. Stop the instance and billing stops with it.

OpenAI-compatible endpoint

Point an existing SDK at your instance by changing the base URL. No rewrite required.

Your weights, your data

Prompts and outputs stay inside your instance. Nothing is logged, sampled or used for training.

Ministral-3 14B Instruct 2512 GPU sizing and cost

What the model needs at each precision, and the cheapest HexGrid Cloud configuration that holds it.

Ministral-3 14B Instruct 2512 vRAM requirement and hourly cost by precision
Precision	vRAM needed	Runs on	Price
BF16 Recommended	34 GB	1× RTX A6000	$0.55/hr
FP8	17 GB	1× RTX 4090	$0.45/hr
INT4	9 GB	1× RTX A4000	$0.20/hr

The official Instruct checkpoint is FP8 and Mistral states it fits in 24GB VRAM. Throughput still depends on GPU, context, batching and vision usage and should be benchmarked on HexGrid.

About Ministral-3 14B Instruct 2512

What Ministral-3 14B Instruct 2512 is built for, and where it falls short.

Ministral 3 14B Instruct 2512 is the largest small-model member of Mistral's Ministral 3 family. It combines a 13.5B language model with a roughly 0.4B vision encoder.

The checkpoint is post-trained for instruction following and distributed in FP8. Mistral explicitly positions the family for edge and private deployment and states that this model can fit within 24 GB of VRAM at FP8.

It supports a 256K context window, image understanding, multilingual prompts, system prompts, native function calling and structured JSON output.

What people run it for

Private AI assistants

Run capable chat and instruction following inside private infrastructure on comparatively modest GPUs.

Local multimodal agents

Combine image understanding, function calling and structured output in local agent workflows.

Document and image analysis

Analyze visual documents, screenshots and images together with long textual context.

Structured automation

Produce JSON and invoke application functions for deterministic application workflows.

Strengths

Fits on a 24 GB GPU in the publisher's FP8 configuration
256K context despite its relatively compact model size
Native image understanding
Native function calling and JSON output
Apache 2.0 licensing
Strong multilingual coverage for an edge-oriented model

Limitations

This is the Instruct checkpoint; Mistral publishes a separate Reasoning checkpoint for workloads requiring its reasoning-specific post-training
Native checkpoint precision is FP8, so comparisons with BF16 model memory footprints require care
Maximum 256K context is expensive to serve even though the weights fit on relatively modest hardware
Image inputs add vision-encoder and context-processing overhead
Publisher benchmark numbers for the Instruct and Reasoning variants should not be mixed

Quickstart

Serve through vLLM and use its OpenAI-compatible chat API.

pythonPOST /v1/chat/completions

from openai import OpenAI

client = OpenAI(
    base_url="https://<your-instance>.hexgrid.cloud/v1",
    api_key="<your-key>",
)

response = client.chat.completions.create(
    model="mistralai/Ministral-3-14B-Instruct-2512",
    messages=[
        {"role": "system", "content": "You are a precise technical assistant."},
        {"role": "user", "content": "Summarize the key risks in this deployment plan."}
    ],
)

print(response.choices[0].message.content)

Mistral's official card demonstrates OpenAI-client usage against vLLM and documents --tool-call-parser mistral for tool use.

Ministral-3 14B Instruct 2512 reported benchmarks

Scores published by the model authors. A starting point, not a guarantee on your own data.

Arena Hard0.551
WildBench68.5
MATH Maj@10.904
MM MTBench8.49

Ministral-3 14B Instruct 2512 specifications

Architecture and serving details for Ministral-3 14B Instruct 2512.

PublisherMistral AIParameters13.9BLicenceApache 2.0ReleasedDecember 2025

Architecture

Language model: 13.5B parameters
Vision encoder: ~0.4B parameters
LM layers: 40
Hidden size: 5120
Attention: 32 heads / 8 KV heads
Checkpoint precision: FP8

Context and modalities

Context: 262,144 tokens
Input: Text + image
Output: Text
Languages: Dozens; 11 explicitly listed on model card

Serving

vLLM: Supported
Function calling: Native
JSON output: Native
Tool parser: mistral
Publisher VRAM claim: Fits in 24 GB at FP8

Ministral-3 14B Instruct 2512 frequently asked questions

Common questions about deploying Ministral-3 14B Instruct 2512 on HexGrid Cloud.

Is Ministral-3 14B actually multimodal?

Yes. It includes an approximately 0.4B-parameter vision encoder in addition to the 13.5B language model.

How much VRAM does it need?

Mistral states that the FP8 Instruct checkpoint can fit in 24 GB of VRAM and requires less when further quantized. Actual available context and concurrency still depend on KV-cache requirements.

Is this the reasoning version?

No. This is the Instruct 2512 checkpoint. Mistral distributes a separate Ministral-3 14B Reasoning 2512 model.

Does it support function calling?

Yes. Mistral lists native function calling and JSON output among its key features.

Deploy Ministral-3 14B Instruct 2512 today

$0.20 per hour on 1× RTX A4000, billed per minute, never shared.

Deploy Ministral-3 14B Instruct 2512 See all models