Try Delta
← All models

Ministral 3

Mistral AI · 2025 · Dense · vision-language

Mistral’s edge family with native vision. Three sizes cover phones through workstations, with a strong cost-to-quality curve for chat, documents, and screenshots, fully offline in Delta.

Maker
Mistral AI
Architecture
Dense · vision-language
Fits from
Phone / edge · 6 GB+ RAM
Context
up to 16K
Runs
Local Intelligence
VisionEdgeMultilingualOn-device chatScreenshotsBalanced hardware

Sizes

Ministral 3 3BQ4_K_M

requires 6 GB+ memory · 16K context · Phone / edge

2.1 GBWeights
Ministral 3 8BQ4_K_M

requires 11 GB+ memory · 16K context · Laptop

5.5 GBWeights
Ministral 3 14BQ4_K_M

requires 17 GB+ memory · 16K context · Desktop

9.5 GBWeights

Memory at context length

Approximate working set for Ministral 3 3B as context grows.

4K
2.6 GB
8K
2.8 GB
16K
2.9 GB

Related families

Gemma 3N

Google’s efficient open model, tuned to run locally on phones, tablets, and laptops.

Gemma 4

Google’s open models for phones and IoT: audio, vision, and near-zero latency on your silicon.

Devstral 2

Mistral’s coding agents for repos, multi-file edits, and long engineering sessions on-device.