Back to all models

Ministral-3 8B Instruct 2512

Deploy Ministral-3 8B Instruct 2512 on a dedicated GPU from $0.20 per hour, billed per minute, on 1× RTX A4000. An edge-oriented multimodal instruction model with 256K context, native function calling and an FP8 checkpoint that Mistral says fits in 12 GB of VRAM.

Multimodal8.8B paramsApache 2.0Deploys in ~5 minSecure Cloud
mistralai/Ministral-3-8B-Instruct-2512

HexGrid Cloud price

$0.20/hr

1× RTX A4000

Context

256K

tokens

Parameters

8.8B

LM + vision

Precision

FP8

Instruct checkpoint

Running Ministral-3 8B 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 8B Instruct 2512 GPU sizing and cost

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

Ministral-3 8B Instruct 2512 vRAM requirement and hourly cost by precision
PrecisionvRAM neededRuns onPrice
BF16
Recommended		22 GB1× RTX 4090$0.45/hr
FP8
11 GB1× RTX A4000$0.20/hr
INT4
6 GB1× RTX A4000$0.20/hr

The publisher provides a 12GB FP8 fit claim, not a universal tokens-per-second figure. Benchmark actual HexGrid throughput under representative text and image workloads.

About Ministral-3 8B Instruct 2512

What Ministral-3 8B Instruct 2512 is built for, and where it falls short.

Ministral 3 8B Instruct 2512 is Mistral's balanced edge-oriented model in the Ministral 3 family. It combines an 8.4B language model with an approximately 0.4B vision encoder.

The instruction checkpoint is distributed in FP8 and Mistral states it can fit in 12 GB of VRAM, making it one of the more practical multimodal models in this group for single-GPU private deployment.

Despite its compact size, it supports 256K context, image understanding, multilingual input, native function calling, JSON output and system-prompt control.

What people run it for

Local daily-driver AI

Chat, summarization and instruction following on a compact private GPU deployment.

Image and document understanding

Describe and reason over screenshots, photos and visually structured documents.

Embedded agents

Use native function calls and JSON output in applications where larger models are impractical.

Multilingual generation

Translation, chat and generation across the dozens of languages supported by the family.

Strengths

  • Publisher states the FP8 checkpoint can fit in 12 GB VRAM
  • 256K context at a compact model scale
  • Image understanding built into the model
  • Native function calling and JSON output
  • Apache 2.0 licence
  • Well suited to private and edge deployments

Limitations

  • Smaller capacity than the 14B model on complex language and reasoning workloads
  • The Instruct checkpoint should not be confused with the separately post-trained Reasoning version
  • A full 256K context still requires significant KV-cache memory even if the FP8 weights fit in 12 GB
  • FP8 requires compatible accelerator/runtime support for native execution
  • Image processing creates additional latency relative to text-only requests

Quickstart

Serve through vLLM and use an OpenAI-compatible chat client.

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-3-8B-Instruct-2512",
    messages=[
        {"role": "user", "content": "Summarize this support ticket and return the next action."}
    ],
)

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

Mistral's official model card demonstrates vLLM serving through an OpenAI client and supports the mistral tool-call parser.

Ministral-3 8B Instruct 2512 reported benchmarks

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

  • Arena Hard0.509
  • WildBench66.8
  • MATH Maj@10.876
  • MM MTBench8.08

Ministral-3 8B Instruct 2512 specifications

Architecture and serving details for Ministral-3 8B Instruct 2512.

PublisherMistral AIParameters8.8BLicenceApache 2.0ReleasedDecember 2025

Architecture

Language model
8.4B parameters
Vision encoder
~0.4B parameters
Checkpoint precision
FP8

Context and modalities

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

Serving

Engine
vLLM supported
Function calling
Native
JSON output
Native
Tool parser
mistral
Publisher VRAM claim
Fits in 12 GB at FP8

Ministral-3 8B Instruct 2512 frequently asked questions

Common questions about deploying Ministral-3 8B Instruct 2512 on HexGrid Cloud.

Can Ministral-3 8B really run in 12 GB VRAM?

Mistral explicitly states that the FP8 checkpoint can fit in 12 GB of VRAM and less when further quantized. Long context and concurrency consume additional memory.

Is it multimodal?

Yes. It contains an approximately 0.4B vision encoder alongside the 8.4B language model.

How large is the context window?

Mistral documents a 256K context window; the checkpoint configuration uses 262,144 positions.

Does it support agents?

Yes. Mistral lists native function calling and JSON output and provides vLLM tool-calling examples.

Deploy Ministral-3 8B Instruct 2512 today

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

Deploy Ministral-3 8B Instruct 2512See all models