Real-World Coding Eval: KAT Coder Outperforms Qwen and Ornith in 35B Local MoE Models

Undici77 · reddit · 2026-08-05

The author conducted months of real-world coding evaluations on 35B MoE local models, focusing on iterative debugging, refactoring, and failure recovery.

Qwen 3.6 (35B A3B): A solid baseline, but tends to hallucinate confidence in edge cases and drifts during longer multi-file reasoning chains.
Ornith 1.0: Excellent benchmarks, but suffers from "overthinking" in practice, leading to high token burn and reasoning loops.
KAT Coder 2.5 Dev: A surprising performer. It offers more decisive outputs with faster convergence, performing best in real-world benchmarks while avoiding the common pitfalls of both Qwen and Ornith.

The author concludes that a similarly sized MoE model with added intelligence and a 1M token context represents a significant step forward.

Original post →

Real-World Coding Eval: KAT Coder Outperforms Qwen and Ornith in 35B Local MoE Models

More from Models