Kat Coder 2.5 is insane. Especially considering I ran it at Q4_K_M
Summary
Kat Coder 2.5 is a coding assistant model that delivers impressive performance even when run at Q4_K_M quantization.
Similar Articles
KAT Coder 2.5 dev: Do yourself a favor and try it!
A developer enthusiastically recommends KAT Coder 2.5 dev, claiming it is faster, more accurate, and uses fewer tokens than Qwen 3.6 35b a3b, and outperforms Gemma 4 models on their setup, with a GitHub repo containing detailed benchmarks.
Kwaipilot/KAT-Coder-V2.5-Dev
KAT-Coder-V2.5-Dev is an open-weight MoE coding model with 35B total parameters (3B active), achieving state-of-the-art results on agentic coding benchmarks through SFT and RL training.
TielCoder's 22 GB 4-bit quant matches Opus4.6 medium on recent real life coding issues, surpassing KAT-Coder and Nail as strongest and fastest MoE picks.
TielCoder is a new 35B-A3B Mixture of Experts model optimized for coding tasks, offering high speed and correctness on real-world issues, surpassing models like Opus4.6 medium and others.
I ran the 35B agentic comparison someone asked for (stock vs Ornith vs KAT-Coder, 120 runs)
A detailed bakeoff of 35B coding models (KAT-Coder-V2.5-Dev, Qwen3.5, Ornith, etc.) with 120 runs shows KAT-Coder matching the best stock pass rate with cleaner tool behavior, while Ornith fails due to mechanical issues. Full methodology and results are linked.
@rasbt: Crazy model! It actually uses the old Qwen2.5-Coder-3B stack and got really great performance with their post-training …
A 3B parameter model using the Qwen2.5-Coder-3B stack achieves coding benchmark scores comparable to Claude Opus 4.5, with detailed post-training techniques including synthetic data, filtering, two-stage SFT, and a novel RL method (MGPO).