My real-world Qwen3-code-next local coding test. So, Is it the next big thing?

A developer conducted a real-world stress test of Alibaba's Qwen3-Code-Next model, running the Q8 MLX version locally on a 128GB Mac Studio Ultra. The ambitious project involved porting KittenTTS-iOS (a Swift application using ONNX and Misaki phoneme libraries) to Windows, representing a medium-difficulty coding challenge with existing code to adapt.

Initial results showed promise as the model successfully handled basic tasks like file operations, web browsing, and system checks. It even managed to build main.cpp, create a JSON parser, and link Windows ONNX libraries, ultimately producing a working WAV file. However, significant limitations emerged as context length increased. The model suffered from frequent client timeouts, inefficient token usage, and struggled with large file edits. At one point, it wasted tokens writing Python scripts to save files and even began editing Swift files incorrectly.

The developer attempted workarounds including adjusting generationConfig.timeout settings and quantizing KV_cache to 8-bit in LM Studio, with mixed results. Despite eventually getting a phoneme test executable to run (processing a 400k phoneme dictionary), the experience highlighted practical deployment challenges. The model's inability to retain learned information between sessions and its tendency to fill context with unnecessary work revealed gaps in its practical utility for complex, multi-step coding projects.