Ministral 3 14B Guide: A Local Model That Sees and Calls Tools

The interesting models for self-hosting are rarely the ones at the top of a leaderboard. They are the ones that fit on hardware you already own, under a licence that lets you ship, while still handling the unglamorous work: classification, extraction, tool calls, a screenshot to describe. Ministral 3 14B is built for exactly that slot.

It is the largest of three small dense models in Mistral's December 2025 release, and the one most likely to replace a 24B model in a local deployment without anyone noticing the swap.

What the 14B is

Mistral announced the Mistral 3 family on 2 December 2025: three dense models at 3B, 8B and 14B parameters under the Ministral name, plus the much larger Mistral Large 3. Every model in the release is Apache 2.0, which permits commercial use, modification and redistribution without a revenue threshold.

The 14B model card breaks the parameter count down as a 13.5B language model plus a 0.4B vision encoder. Key specifications from Mistral's documentation:

• Context window — 256K tokens
• Input — text and images, with near-square images recommended
• Capabilities — native function calling, JSON output, structured outputs, system prompt adherence
• Languages — English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean and Arabic among others

Mistral prices the hosted model at $0.20 per million tokens on both input and output, and OpenRouter listed the same at the time of writing. Flat pricing like that is unusual and makes output-heavy work cheaper than on most models where output costs several times input.

Fitting it on one card

The instruct model ships with FP8 weights, and Mistral says it can be deployed locally in about 24GB of VRAM at that precision, with less if quantised further. That puts it within reach of a single 24GB workstation or high-end consumer card, which is the line that separates "we can run this on a spare machine" from "we need to provision servers".

The 256K window comes with a caveat that applies to every local model. The weights fitting is not the same as long prompts fitting. The KV cache grows with context, so a full 256K prompt needs far more memory than the 24GB figure, and on a single card you will cap context well below the maximum. Decide what context you actually need and size for that. Quantization explained covers the precision trade if you want to go lower than FP8.

Base, instruct and reasoning are different products

Each Ministral 3 size comes in three variants, and the published numbers belong to specific ones. Mixing them up is the most common way to misread this model.

• Instruct is the general model, and the one most hosted endpoints serve. The model card reports Arena Hard 0.551 against 0.427 for Qwen3 14B in non-thinking mode and 0.436 for Gemma 3 12B, and WildBench 68.5 against 65.1 and 63.2.
• Reasoning trades latency and tokens for depth. Mistral reports 85.0% on AIME 2025 for it, against 73.7% for Qwen3 14B in thinking mode, along with GPQA Diamond 71.2% and LiveCodeBench 64.6%.
• Base is the pre-trained model for your own fine-tuning.

If you read 85% on AIME and then call the instruct endpoint, you are not getting that model. Mistral also claims the instruct variants often produce an order of magnitude fewer tokens than comparable models, which, if it holds on your prompts, matters as much as the per-token price.

14B, 8B or Mistral Small 3.2

Mistral describes the 14B as offering performance comparable to Mistral Small 3.2 24B, the next model up. That gives you a ladder to choose from:

• Pick the 14B when you want Small-class quality in roughly half the memory, or when vision, tool calling and the Apache licence all need to come together on one GPU.
• Drop to the 8B for high-volume classification, routing or extraction, where latency matters more than the last few points of quality. The Ministral 3 8B guide covers it, and the 3B guide covers the edge and on-device case.
• Move up to Small 3.2 if your tests show the 14B failing on instruction-heavy prompts and you have the memory. The Mistral Small 3.2 guide covers that model.

For agentic coding over a repository, none of these is the right tool. A model built for long tool loops will do better, and the 14B's strengths lie in bounded tasks.

Testing it for a local deployment

Mistral's model card makes two recommendations worth following: a temperature below 0.1 for production use, and small, well-defined tool sets rather than long lists. Small models degrade faster than large ones as the tool list grows.

A practical test plan:

1. Take fifty real inputs from the job you want to move, with known correct outputs.
2. Run them through your current model and the 14B instruct at the same low temperature.
3. Score exact-match for extraction and classification, and schema validity for JSON output.
4. For tool calling, count wrong-tool and malformed-argument errors separately; they have different fixes.
5. Measure tokens per answer. If the efficiency claim holds, your effective cost drops further than the price suggests.

If the 14B matches your current model on that set, you have a smaller, permissively licensed model that runs on one card. That is usually worth more than a few points on a benchmark you do not use.