Hands-on with Cursor Max Mode: Building a Mistral Small 3.1 Multimodal App for $1.7 (With Source Code)

Hands-on with Cursor Max Mode: Building a Mistral Small 3.1 Multimodal App for $1.7 (With Source Code)

Today, I’m excited to introduce Sonnet Max mode in the latest Cursor release and share my hands-on experience using this powerful feature.

For just $1.7 and two quick calls to Sonnet thinking mode, I successfully built a Mistral Small 3.1 API-powered multimodal application. This app not only performs image analysis and text-based conversations but also includes a convenient chat history review feature.

1. App Demo: Intuitive and User-Friendly

Let’s first take a look at the application interface built with Cursor Max mode:

As you can see, the app features a clean, modern UI with straightforward functionality. Users can easily:
✅ Upload images for analysis
✅ Engage in text conversations with the model
✅ Review past interactions via the history button

At its core, the app leverages Mistral Small 3.1, which provides robust multimodal capabilities.

2. Mistral Small 3.1: Small Model, Big Potential

Mistral Small 3.1 is the latest model from Mistral AI, an upgrade over Small 3 with enhanced multimodal understanding.

Key Features:
🔹 Multimodal Comprehension – Processes both text and images
🔹 Extended Context (128K) – Handles longer conversations and complex queries
🔹 High Performance (24B Parameters) – Outperforms Gemma 3 27B and GPT-4o Mini in many benchmarks

Performance Comparison (Mistral Official Data):

Limitations:

Math performance slightly lags behind Gemma 3 27B
Optimized for English & French – East Asian language support still needs improvement

3. Cursor Max Mode: The Ultimate Development Tool

Cursor Max mode is the smartest, longest-context, and most “thoughtful” mode available in Cursor.

Cost: $0.05 per request & per tool use
First-time setup: Enable “Usage-based pricing” in settings
Pro Tip: Set a “Hard Limit” to control costs

4. Step-by-Step Development: From Prototype to Polish

Preparation:

Created a new app.py file in Cursor
Copied Mistral’s official Python code example into app.py
Modified the model name to mistral-small-latest

Initial Build:

Activated Agent Mode → Max Mode
Entered the prompt:

Cursor auto-generated the app code, including necessary files and folders.

First request cost: ~$0.6 (producing a functional MVP)

Optimization & Refinements:
⚠ Caution: The 3.7 thinking mode sometimes adds unnecessary restrictions—always review generated code carefully.

Key Improvements:
🔸 Added support for multiple image sizes/formats
🔸 Fixed Base64-encoded image display in chat history
🔸 Enhanced UI/UX design for better usability

5. Multimodal Testing: Strong but Room for Growth

I tested Mistral Small 3.1’s capabilities with various inputs:

✅ IKEA Sofa Assembly Diagram

Recognized most parts but misidentified washers as screws

✅ Chart Analysis

Accurately interpreted trends, peaks, and volatility

⚠ Chinese Comic Recognition

Mistranslated “轻音乐” (light music) as “can’t hear”
Needs improvement in Chinese OCR

✅ English Chart Recognition

Provided precise metric estimations

Verdict: Excellent with English/Western content, but East Asian language support needs work.

Source Code Available!

Want to try this app yourself? Grab the code here:
🔗 https://github.com/nicekate/mistral-flask-app

Final Thoughts:
For just $1.7, Cursor Max mode + Mistral Small 3.1 delivers impressive multimodal functionality. While not perfect (especially for non-Latin scripts), it’s a cost-effective way to prototype AI-powered apps.

What will you build with it? 🚀