I Built the Same App With MiniMax M3 and Composer
Tech With Tim ·2026-06-22 ·2 min read
Summary written by us from the video's transcript. The video, and everything in it, is Tech With Tim's work.
Learn how to set up MiniMax M3 in Cursor, compare its coding performance and cost against Composer 2.5 across several tasks, and understand the trade‑offs in speed, output quality, and token pricing.
Takeaways
- MiniMax M3 can be integrated into Cursor by overriding the OpenAI base URL and adding a custom model name.
- Its token pricing provides hundreds of times more tokens per dollar than comparable models like Opus or Claude.
- For complex tasks, MiniMax generates more structured code, tests, and validation but takes 5–6× longer than Composer.
- Composer is faster for simple scaffolding but often omits robustness features such as testing and input validation.
- The large context window (up to 1 million tokens) enables MiniMax to handle extensive codebases and multi‑step reasoning.
What is MiniMax M3?
MiniMax is a Chinese lab that releases open weights for its models, allowing download, fine‑tuning, and offline use—unlike closed models such as Anthropic or GPT. The M3 model supports up to a 1 million token context window (minimum 512 k), multimodal inputs (image, video, text), and can be run via API or locally. It launched on June 1 and is positioned as a cost‑effective alternative for coding tasks.
Pricing Advantage
MiniMax offers token plans that deliver far more tokens per dollar than major competitors: $20 buys ~1.7 billion MiniMax tokens versus 11 million Claude Haiku, 3.7 million Claude Sonnet 4.6, or 2.2 million Opus 4.8 tokens. This translates to roughly 765× the token volume of Opus for the same spend, making it attractive for heavy‑use scenarios.
Configuring MiniMax M3 in Cursor
1. Obtain a MiniMax subscription key from the dashboard. 2. In Cursor settings, enable the OpenAI API key field and override the base URL with the appropriate regional endpoint (e.g., https://api.minimax.cn/v1). 3. Add a custom model named **minimax‑M3** via *View all models → Add custom model*. 4. Save and test with a simple prompt like “hello world” to confirm connectivity.
Task 1: Building a URL Shortener
Both MiniMax M3 and Composer 2.5 were given the same prompt to create a Node.js URL shortener with an API endpoint, dashboard, and modern styling. MiniMax produced 21 well‑structured files, full test coverage, validation logic, and dark/light mode support; it took ~15 minutes and used about 106 k tokens. Composer generated a simpler 7‑file solution without tests or validation in roughly 2 minutes. The trade‑off: MiniMax delivered richer, more maintainable code at the cost of longer runtime.
Task 2: Rust Ray Tracer
The prompt asked each model to write a Rust ray tracer that renders 4–5 spheres on a checkered ground to PNG. MiniMax’s output rendered a correctly oriented scene with decent dimensionality, while Composer’s result was upside‑down and had visual artifacts. Both models took similar code size, but MiniMax required 3–4× more processing time, reflecting deeper analysis of the task.
Task 3: Large Refactor – Adding Daily Streak Feature
A complex existing Python/TypeScript codebase was used to test how each model adds a new “Daily Streak” feature. Composer finished quickly but made minimal changes and added only one small UI component. MiniMax took longer, performed many tool calls, added more detailed code, comments, and a functional streak UI that displayed best‑streak data. Compilation succeeded for both, but MiniMax’s output appeared more maintainable.
Overall Observations
MiniMax M3 is not the fastest model; Composer consistently completed tasks in a fraction of the time. However, MiniMax leverages its large context window to perform extensive analysis, generate comprehensive code structures, and include testing/validation—resulting in higher‑quality outputs for most tasks. Token usage remains well within limits (under 2 % of a five‑hour quota), reinforcing its cost‑effectiveness.