Fine-tuned from Qwen/Qwen3-32B with QLoRA on SWE (software engineering)
problem-solving agent traces. The model follows a THOUGHT + single bash command
format in multi-turn shell interactions.
Intended use
Host as an LLM API (Hugging Face Inference Endpoints, TGI, vLLM).
Agentic coding tasks requiring step-by-step shell commands and reasoning.
