Workflows/Build a local AI agent with Ministral 3 Reasoning

Agents / WORKFLOW

Build a local AI agent with Ministral 3 Reasoning

Run an official Ministral 3 Reasoning GGUF on your computer, allow one bounded local lookup, and inspect the tool result and final answer.

EstimatedSource checked 9/29/2026

LOCALRENTED GPU

Freshness 26 out of 100. Repository age: 349 days old, Oct 31, 2025. Measures age, not quality.

Popularity 65 out of 100. 73,973 downloads in 30 days; 167 lifetime likes. Checked Oct 15, 2026. Measures adoption, not quality.

What do these scores mean?

Two ways to compare models. Neither score measures answer quality, benchmark performance, or whether a model fits your machine.

Freshness · release age

A new release starts at 100. The score halves every 180 days: 100 × 2^(−age in days / 180). When the release date is unavailable, a repository date is a fallback—not proof of when the model was released. Missing dates stay unscored.

Check the date source

Popularity · adoption

Downloads over the past 30 days carry 80% of the score; lifetime likes carry 20%. Both use a logarithmic scale, capped at 10 million downloads and 100,000 likes. If likes are unavailable, downloads carry all the weight. Missing download statistics stay unscored.

round(80 × min(1, log10(1 + downloads) / 7) + 20 × min(1, log10(1 + likes) / 5))

Check the statistics source

Scores as of Oct 15, 2026. Statistics checked Oct 15, 2026. Statistics are marked for refresh after 72 hours.Full scoring method

Before you begin

Hardware

Estimated memory fit for the three publisher Q4_K_M GGUF packages with one tool and 4K context. Actual speed and GPU offload depend on the runtime and other apps.

Machine	Ministral 3 · 3B Reasoning Q4_K_M	Ministral 3 · 8B Reasoning Q4_K_M	Ministral 3 · 14B Reasoning Q4_K_M
NVIDIA GPU · 8 GB	Runs the small 3B task with headroom.	May use CPU offload; 12 GB is a simpler start.	Likely needs CPU offload; choose 16 GB for full GPU load.
NVIDIA GPU · 12 GB	Runs with headroom.	Runs with room for the small tool loop.	Estimated fit at 4K; limited headroom.
NVIDIA GPU · 16 GB	Runs with headroom.	Runs with headroom.	Runs the 14B Q4 package at 4K with more room.
AMD Radeon · 16 GB	Runs if Ollama GPU backend works; check ollama ps.	Runs if Ollama GPU backend works; check ollama ps.	Estimated 4K fit if Ollama GPU backend works.
Intel Arc · 16 GB	Runs if Ollama Vulkan backend works; check ollama ps.	Runs if Ollama Vulkan backend works; check ollama ps.	Estimated 4K fit if Ollama Vulkan backend works.
Apple Silicon Mac · 8 GB unified	Estimated 4K starting point; close other apps.	Needs more memory; use a 16 GB Mac.	Needs more memory; use a 16 GB-plus Mac.
Apple Silicon Mac · 16 GB unified	Runs with headroom.	Estimated 4K fit; limited headroom.	Estimated 4K fit; limited headroom.
Apple Silicon Mac · 24 GB unified	Runs with headroom.	Runs with headroom.	Runs with headroom at 4K.
CPU-only · 16 GB RAM	Can run in RAM; slower than GPU.	May need more RAM as context grows.	Use 24 GB-plus RAM or a smaller tier.
CPU-only · 32 GB RAM	Can run in RAM; slower than GPU.	Can run in RAM; slower than GPU.	Can run in RAM; slower than GPU.

Memory estimates are not performance measurements. Tool-use reliability should be checked with the exact included test on each selected quantization.

LOCAL AGENT / MINISTRAL 3SOURCE-LINKED SETUP

Run one safe local tool with Ministral 3 Reasoning

Choose the 3B, 8B or 14B publisher GGUF for your machine. Ollama pulls the exact Reasoning 2512 Q4_K_M variant; a short Python script allows one local lookup and checks the complete tool round trip on Windows, Mac or Linux.

Choose a package Run its commands Check the result

BEFORE YOU BEGIN

Choose your package

Select the weights you want. The instructions, downloads and settings change with your choice.

Save your machine in My Hardware to get an automatic starting choice. You can always choose any package yourself.

CURRENT SETUP

Ministral 3 · 3B Reasoning Q4_K_M · official

The 2.15 GB Q4_K_M weight file needs extra runtime and KV memory. Start with 4K context. Estimate 6 GB GPU or 8 GB Apple Silicon unified memory; 8 GB GPU gives more room. Check ollama ps for GPU/CPU placement.

Freshness 26 out of 100. Repository age: 349 days old, Oct 31, 2025. Measures age, not quality.

Popularity 48 out of 100. 4,030 downloads in 30 days; 52 lifetime likes. Checked Oct 15, 2026. Measures adoption, not quality.

What do these scores mean?

Two ways to compare models. Neither score measures answer quality, benchmark performance, or whether a model fits your machine.

Freshness · release age

Check the date source

Popularity · adoption

round(80 × min(1, log10(1 + downloads) / 7) + 20 × min(1, log10(1 + likes) / 5))

Check the statistics source

Scores as of Oct 15, 2026. Statistics checked Oct 15, 2026. Statistics are marked for refresh after 72 hours.Full scoring method

View weight source

THE WORKFLOW

Set up Ministral 3 · 3B Reasoning Q4_K_M on your machine

Choose where you want to run it. The instructions below update to match your selection.

CHOOSE YOUR SETUPMac

Use Terminal on Apple Silicon. Ollama for macOS is available from the official download page.

Install Ollama on Mac

Install the official Mac app, open it, then start Terminal.

Official Ollama download

Check the runtime and Python

In Terminal, confirm Ollama and Python 3 are available. Install Python 3 first if python3 is not found.

Install Python 3

ollama --version
python3 --version

Download the exact 3B Reasoning GGUF

Pull the exact package before running the example, so model download failures are separate from Python errors.

ollama pull hf.co/mistralai/Ministral-3-3B-Reasoning-2512-GGUF:Q4_K_M

Install the local Python libraries

The official Ollama Python client connects to the service on your own machine.

python3 -m pip install --upgrade ollama

Create tool_agent.py

In Terminal, paste this block. It writes the complete small example as tool_agent.py in your current folder. Read the code before running it. This package uses the exact publisher 3B Reasoning GGUF; no model edit is needed after choosing it.

cat > tool_agent.py <<'PY'
from ollama import chat

def get_temperature(city: str) -> str:
    """Return the temperature from this local example lookup, not live weather."""
    return {"New York": "22°C", "London": "15°C"}.get(city, "No sample value")

messages = [{"role": "user", "content": "Using your tool, what is the sample temperature in New York?"}]
response = chat(model="hf.co/mistralai/Ministral-3-3B-Reasoning-2512-GGUF:Q4_K_M", messages=messages, tools=[get_temperature], think=True)
messages.append(response.message)
if not response.message.tool_calls:
    raise RuntimeError("No tool call: check that hf.co/mistralai/Ministral-3-3B-Reasoning-2512-GGUF:Q4_K_M loaded and retry the exact question")
call = response.message.tool_calls[0]
if call.function.name != "get_temperature" or call.function.arguments.get("city") != "New York":
    raise ValueError("Unapproved tool name or city; no function was executed")
result = get_temperature(**call.function.arguments)
print("Local tool result:", result)
messages.append({"role": "tool", "tool_name": "get_temperature", "content": result})
final = chat(model="hf.co/mistralai/Ministral-3-3B-Reasoning-2512-GGUF:Q4_K_M", messages=messages, tools=[get_temperature], think=True)
print("Final answer:", final.message.content)
PY

Execute the tool loop

Run the Python script. It checks that the requested function is exactly get_temperature and the city is New York, executes the local lookup, returns a tool message, then asks Ministral 3 3B Reasoning for its final answer.

python3 tool_agent.py

Check the local result and placement

Look for Local tool result: 22°C and a final answer that uses 22°C. If the model does not request the tool, the script raises an error rather than pretending it succeeded. Use ollama ps to see CPU/GPU placement.

ollama ps

SUCCESS CHECK · Ministral 3 · 3B Reasoning Q4_K_M

A complete agent turn with a checked tool result

The script prints Local tool result: 22°C and a final answer that uses it. It rejects any unexpected function name or city before executing the lookup; no shell or web tool is granted.

These are publisher GGUFs of the Reasoning 2512 checkpoints, not the separate Ministral Instruct model. Use a short 4K starting context; the publisher maximum of 256K is not a practical default on small machines. The lookup is static demonstration data, not live weather.

Back to setup steps

When it doesn’t go to plan

The local script reports no tool call.

Confirm the exact GGUF Ollama tag is installed and rerun the sample question. The script stops rather than inventing a tool result; smaller quantizations may be less reliable at tool use.

Ollama reports insufficient memory.

Choose the 8B or 3B package, close other GPU tasks and keep the first test at short context. Use ollama ps to check whether the model is spilling into system RAM.

The final response ignores the lookup value.

Confirm the first assistant tool-call message and the role="tool" result are appended before the second chat request. Keep only the named get_temperature tool available.

REFERENCE LIBRARY

Sources and files

Original instructions, model files, and compatibility notes behind this setup.

5 SOURCES

01Exact publisher 14B Reasoning GGUF and Ollama launchhuggingface.co 02Publisher 8B Reasoning GGUFhuggingface.co 03Publisher 3B Reasoning GGUFhuggingface.co 04Ollama tool-calling API sequencedocs.ollama.com 05Ollama download and GPU supportollama.com

The model behind this workflow

Ministral 3 14B Reasoning 2512

Keep exploring

Text

SETUP GUIDE

Qwen3.5 · 9B private chat setup

Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.

Freshness 41 out of 100. Release age: 230 days old, Feb 27, 2026. Measures age, not quality.

Build a private coding assistant with Qwen3.8 27B

Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.

Freshness 76 out of 100. Repository age: 71 days old, Aug 5, 2026. Measures age, not quality.

Arena Open Source Web coding: rank 10 of 55. Ranked model: qwen3.8-27b. Original model ranking, not a benchmark of this package. Export dated Oct 13, 2026. Source checked Oct 15, 2026. Human preference ranking, not hardware suitability.+3

SOFTWAREOllama + Visual Studio Code + Cline + Python

Create images locally with FLUX.2 Klein

Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.

Freshness 35 out of 100. Release age: 273 days old, Jan 15, 2026. Measures age, not quality.

SOFTWAREComfyUI

LOCALRENTED GPU

Verified source