Ministral 3 14B
Ministral 3 14B is the largest of Mistral's December 2025 dense edge models. It combines a 13.5B language model with a 0.4B vision encoder. The hosted instruction model supports structured responses and tool calls. Its weights are available under Apache 2.0.
Choose the checkpoint and precision
The official Instruct checkpoint uses FP8. Mistral states that it can fit in 24GB of VRAM. The family also has separate Base and Reasoning checkpoints, so a download marked Base is not the same assistant as the hosted instruction model.Ministral 3 14B Instruct model card
Use the instruction API
The hosted model identifier is ministral-14b-2512. Its model page lists function calling, structured output, document questions, prefix completion and batch processing. The Base and Reasoning names here identify published weight variants; they are not additional confirmed serverless aliases.Ministral 3 14B API model page
Check the device workload
Practical guidance. This size is useful to evaluate when a local assistant must read images and follow multi-step instructions, and smaller models fail those tasks. Leave room for the image input, context cache and concurrent requests. A checkpoint fitting in memory does not establish full-context serving capacity.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable |
| Model identifiers | ministral-14b-2512Ministral-3-14B-Instruct-2512Ministral-3-14B-Base-2512Ministral-3-14B-Reasoning-2512 |
| Context window | 262,144 tokens The official Instruct serving example sets a maximum model length of 262144; available capacity depends on deployment memory. Ministral 3 14B Instruct model card |
| Release | 2 December 2025; general availabilityMinistral 3 14B API model page |
| Components | 13.5B language model and 0.4B vision encoderMinistral 3 14B Instruct model card |
| License | Apache 2.0Ministral 3 14B API model page |
Related models
- Mistral Small 4
Mistral AI · Mistral Small
An Apache-licensed multimodal model that combines instruction following, reasoning and coding with selectable reasoning modes.
Launch date not confirmedAPI · Downloadable - Ministral 3 3B
Mistral AI · Ministral 3
The smallest Ministral 3 release pairs a 3.4B language model with a 0.4B vision encoder and offers multiple training variants.
Launch date not confirmedAPI · Downloadable - Ministral 3 8B
Mistral AI · Ministral 3
An 8B multimodal Ministral release with separate pretrained, instruction and reasoning checkpoints for local deployment.
Launch date not confirmedAPI · Downloadable