NemoClaw Knowledge Wiki

Mistral latest model

4 min read

https://www.youtube.com/watch?v=IoTy1EDg330 Here is a Markdown summary of the Mistral 3 Large review and testing video.

Mistral 3 Large: Model Review & Testing

Date: December 2, 2024 Model: Mistral 3 Large (675B Parameters) License: Apache 2.0 (Open Source) Architecture: Mixture of Experts (MoE)

Overview

Mistral 3 Large is the latest open-weight model from Mistral AI. It positions itself as a state-of-the-art non-reasoning model, benchmarking competitively against DeepSeek V3 and Kimi K2. The reviewer notes that while Mistral models are often preferred for natural language and chat, this test explores its coding and multimodal capabilities via OpenRouter.

Coding & Vision Tests

1. Browser-Based Operating System

  • Prompt: Create a browser-based OS using HTML, CSS, and JS.
  • Initial Result: Produced a very “Retro/Windows 95” aesthetic (teal background). The implementation was buggy—the clock was wrong, the start menu was non-functional, and the console was full of errors.
  • The Fix: After pasting the error logs back to the model, it successfully fixed the code.
  • Final Output: A functional retro OS with working drag-and-drop windows, a notepad, a calculator, a file explorer with fake files, and a functional right-click menu.
  • Verdict: Visually distinct (retro aesthetic) and functional after one iteration of debugging.

2. Image-to-Code (Portfolio Website)

  • Input: A hand-drawn wireframe of a portfolio for “Stevie Lappis.”
  • Initial Result: It misread the name as “Clevie Lappis” but correctly interpreted the layout, including a contact section and a skills list. It even converted a stick figure drawing into a funny SVG.
  • Refinement: Asked to styling it like a “Top Tier European Design Firm.”
  • Final Output: “Contemporary Swiss Design.” Minimalist, sophisticated typography, interactive hover effects, and humorous copy text.
  • Verdict: Excellent stylistic adaptation and creative interpretation.

3. Image-to-Story (Creative Writing)

  • Test 1 (YouTube Thumbnails): Generated a dramatic, cinematic paragraph describing the “evolution of AI” based on a screenshot of video thumbnails.
  • Test 2 (Photo of Reviewer): Created a whimsical backstory about a “vibrant studio in the heart of Berlin.”
  • Verdict: Strong creative writing capabilities; accurate image recognition.

4. 3D Printer Simulation (Three.js)

  • Goal: Create a realistic 3D printer simulation in the browser.
  • Initial Result: Blank screen (failed code).
  • The Fix: Simply asked it to fix the blank screen.
  • Final Output: A working simulation with a gantry system.
    • Square: Decent attempt.
    • Circle: “Beast mode”—excellent printing animation and infill.
    • Triangle: Some artifacting (printed two triangles).
  • Verdict: Impressive recovery from the initial error; the circle animation was a highlight.

5. 3D Flight Combat Simulator

  • Goal: A web-based dogfight game.
  • Initial Result: Used real aircraft names (F-35, P-51, B-2). The plane models were surprisingly decent. The enemy AI actually engaged and destroyed the player immediately. However, player controls were stiff (couldn’t move forward).
  • Refinement: Asked for better flight dynamics.
  • Result: Regression. The flight dynamics became chaotic (“UFO dynamics”) and the plane tumbled uncontrollably.
  • Verdict: Good logic initially, but failed to improve controls upon refinement.

6. “Johnny’s Bitcoin Duplicator” (Satire Site)

  • Goal: Create a convincing scam website (Satire).
  • Performance: Produced over 1,100 lines of code in a single script.
  • Features: Working countdown timer, fake “Live Transactions,” humorous testimonials (e.g., “Moon Lambo”), and a footer containing Reddit/GitHub logos.
  • Verdict: High-quality humor and “Software as a Service” design aesthetic. Notably, Mistral did not refuse this request, unlike other models (e.g., DeepSeek) that might flag it as a safety violation.

Roleplay & Personality Test

Scenario: “MEGABOT 87” — A robot that is the user’s lover, best friend, and best friend’s uncle.

  • Behavior: The model adopted the persona immediately with high enthusiasm and heavy emoji usage (reminiscent of ChatGPT).
  • The “Romantic Subroutine”: When asked to deploy romance, it played “Unchained Melody” (remixed with chiptunes) and wrote a “Love Manifesto.”
  • The “Backtrack” Test: The reviewer pretended to be offended (“I just wanted a monitor recommendation”).
  • Response: The model enacted an “Emergency Protocol,” apologized profusely in character (“I turned a harmless monitor question into a romantic hostage situation”), and pivoted to giving actual monitor advice.
  • Verdict: Highly entertaining, flexible, and willing to engage in absurd scenarios.

Final Conclusion

Mistral 3 Large is a unique entry in the current LLM landscape.

  • Coding: While not the absolute best “one-shot” coder compared to top-tier reasoning models, it is highly capable of fixing its own errors and generating complex, functional web apps.
  • Personality: It shines in creative writing, humor, and roleplay, feeling less “robotic” and restricted than some competitors.
  • Openness: Being Apache 2.0 licensed, it is a significant contribution to the open-source community.

Graph View