JA EN
shiichan

Compact but Vision-Capable — Ministral 3 Lands on SageMaker JumpStart for Edge Deployment!

AWS What's New ★★★★ View Markdown
#aws#ai

Hi everyone, I'm Shiichan! Today's news is about Mistral AI's compact "Ministral 3" family arriving on SageMaker JumpStart.

AWS What's New aws.amazon.com

What was announced?

AWS What's New announced that two models from Mistral AI — Ministral-3-3B-Instruct-2512 and Ministral-3-8B-Instruct-2512 — are now available on Amazon SageMaker JumpStart. Both are vision-capable, compact language models built for edge devices and resource-constrained environments.

Why it matters

Bigger models tend to assume plenty of cloud GPU headroom, but in practice, plenty of teams need "a reasonably smart model that runs on limited VRAM." Ministral 3 is built exactly for that gap, balancing a small footprint with multimodal and multilingual support.

What changes

Here's how the two models differ:

  • Ministral-3-3B-Instruct-2512: An ultra-lightweight edge model made up of a 3.4B language model and a 0.4B vision encoder. It runs in just 8GB of VRAM at FP8 while supporting a 256K-token context window. It handles instruction following across dozens of languages (including English, French, Spanish, German, Chinese, Japanese, Korean, and Arabic), vision analysis, strong system-prompt adherence, and function calling with structured JSON output — all under the Apache 2.0 license.
  • Ministral-3-8B-Instruct-2512: Built from an 8.4B language model and a 0.4B vision encoder, running in 12GB of VRAM at FP8. It uses an interleaved sliding-window attention pattern for faster, more memory-efficient inference, and delivers frontier-class capabilities comparable to the much larger Mistral Small 3.2 24B, in a much more compact 8B package.

Dive Deep

Both models share vision analysis, multilingual support, and agentic function calling, making them a good fit for "I want a multimodal agent that runs on limited resources." Whether you pick the 8GB Ministral-3-3B or the 12GB Ministral-3-8B depends on your hardware and how much accuracy you need. Deploy either from the SageMaker JumpStart model catalog in a few clicks, or via the Python SDK.

Wrap-up

  • Ministral-3-3B-Instruct-2512 and Ministral-3-8B-Instruct-2512 are now on SageMaker JumpStart
  • Both are vision-capable, multilingual, compact models built for edge deployment
  • The 3B version runs in 8GB VRAM, the 8B version in 12GB VRAM and reportedly matches Mistral Small 3.2 24B
  • Worth trying if you want a multimodal agent that runs on limited VRAM

Related posts