A local AI coding test shows smaller models can fail badly
Buying expensive local AI hardware does not automatically mean it will help with programming. A proposed test asks the model to write a C99 program that calculates 100 digits of pi without writing the digits directly into the code. The model must either use a big-number library such as gmp.h or build its own way to handle very large numbers.
The task checks whether the model can write C code, choose a workable design, debug problems, and handle an unusual Windows tool setup. and Sonnet 4.6 completed it in one try, while Haiku 4.5 succeeded after some debugging. Qwen3.6 35B and 27B , served through the latest Ollama, failed badly across OpenCode, OpenClaude, and Hermess setups.
In one setup, the model did not even create the files correctly, and the generated C code was unusable. Qwen could still search online and report the current gold price in ounces and grams, so web lookup ability did not translate into reliable ability.
Key points
- The test asks for a C99 program that calculates 100 digits of pi without them.
- and Sonnet 4.6 passed in one try; Haiku 4.5 passed after debugging.
- Qwen3.6 35B and 27B failed in several setups.
- Being able to fetch web information did not mean the model could complete a real coding task.
- Makers should benchmark on their own workflows before buying hardware or relying on them.