QClaw-4B matches giant models on agentic benchmark

QClaw-4B demonstrates that specialized fine-tuning on task trajectories can act as a massive force multiplier for agentic performance, allowing a 4-billion parameter model to rival much larger generalists. It achieves an 84.8 Claw Score, effectively tying with frontier models such as Kimi K2.5 and GLM-4.5 on the ClawBench benchmark. Built on the Qwen3.5-4B architecture, the model is specifically optimized for OpenClaw frameworks and represents a significant validation of the "smol" model trend in local agent development.