Can I run Ministral 3 14B?

Ministral 3 14B is Mistral's December 2025 mid-size model, a 14B-parameter dense network that also handles vision, so it's the pick when you want one local model that chats and reads images without juggling two checkpoints. At a 4-bit quant it lands around 8.5 GB, just past what an 8 GB card holds comfortably but right at home on a 12 GB RTX 3060 or any Apple Silicon Mac with 16 GB of unified memory or more. Drop to a 2-bit build and it shrinks to about 5.9 GB if you're squeezed; the q8 build is closer to 15 GB.

In daily use it feels quick for its size. On an RTX 3060 12 GB you'll see around 36 tokens per second at 4-bit, comfortably faster than reading speed, and an RTX 4090 pushes that to roughly 101 tok/s. An M-series Max sits near 41 tok/s. The 256K context window is the headline number, but treat it as a ceiling, not a default: even at 128K, total memory climbs to about 30 GB once the KV cache fills, well beyond a single 12 GB or 16 GB device. Keep working context to a few thousand tokens unless you have a 24 GB-plus card to spare.

Against its neighbours, Ministral 3 14B generally has the edge over older siblings like Mistral 7B and Mistral Nemo 12B on harder, multi-step prompts simply by being larger and newer, while OLMo 2 13B is the fully-open-data alternative if reproducibility matters more to you than raw capability. Its real standout is the built-in vision capability at this size with a genuinely usable speed profile. And the license is the easy part: Apache 2.0, so you can use it commercially and in production with no provider-specific strings attached.