Jackrong/Qwopus3.5-9B-Coder-GGUF · Hugging Face

Qwopus3.5-9B-coder is a community fine-tuned 9B parameter model built on Qwen3.5-9B, optimized for agentic coding, complex tool calling, and logical reasoning. It uses a novel Trace Inversion technique to reconstruct full reasoning chains from compressed 'reasoning bubbles' of commercial models like Claude, combined with real GLM-5.1 agent trajectory data. Training follows a three-stage curriculum learning pipeline. The model runs efficiently at 8-bit precision on 16GB RAM devices and supports GGUF quantization formats compatible with llama.cpp, Ollama, LM Studio, vLLM, and other local inference tools. It also supports vision and tool calling. Benchmark results on HermesAgent-20, ToolCall-15, and BugFind-15 are provided, along with guidance for extending context beyond 32K using RoPE/YaRN scaling.