Jackrong Qwopus‑3.6‑27B‑Coder: 27‑B Agentic Coding Model (GGUF)
Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF ↗
Qwopus‑3.6‑27B‑Coder is a 27‑billion‑parameter LLM released in GGUF format for efficient inference with llama.cpp and llama‑cpp‑compatible runtimes. Billed as a “Coder SFT” release, it builds on the reasoning‑rich Qwopus‑3.6‑27B‑v2 base model and adds a focused fine‑tuning stage that targets agentic coding, tool‑use stability, code debugging, and repository‑level tasks. The model is trained with the Unsloth framework, which provides memory‑optimized large‑model fine‑tuning, and benefits from a close hardware partnership with engineer Kyle Hessling. It is licensed under the Apache‑2.0 license and distributed as a GGUF checkpoint, making it directly usable with llama.cpp, gpt‑cpp, and other spec‑compatible inference engines.
The fine‑tuning data emphasizes “Trace Inversion” – a process that reconstructs compressed reasoning bubbles into full step‑by‑step chain‑of‑thought (CoT) traces – and high‑quality multi‑turn agent trajectories that include explicit tool definitions, tool calls, and real feedback. Three special datasets are used: Jackrong/claude‑opus‑4.6‑traceInversion‑9000x, Jackrong/claude‑opus‑4.7‑traceInversion‑5000x, and lambda/hermes‑agent‑reasoning‑traces, together providing thousands of detailed coding and tool‑use examples. The model inherits a native 32K token context window (with RoPE/YaRN scaling for longer contexts) and supports speculative decoding via a Multi‑Token Prediction (MTP) variant that offers ~1.66× speed‑up.
Performance-wise, the first published benchmark is SWE‑bench Verified (full 500) run in a no‑thinking mode, where the model solved 335 of 500 tasks (≈66.9% pass rate) without relying on visible reasoning traces. On a high‑end GPU (e.g., RTX 5090) with MTP enabled, it runs at roughly 100 tokens/sec, making it suitable for interactive local coding agents. The model is released under the Apache‑2.0 license, includes the MTP variant for speculative decoding, and is fully compatible with the llama.cpp ecosystem and other GGUF‑compatible runtimes.
Project Ideas
- Build a local AI coding assistant that can generate, debug, and refactor code across an entire Git repository using tool‑calling for file operations.
- Create an image‑to‑code utility that takes screenshots of code snippets or UI mockups and returns executable code, leveraging the model's image‑text‑to‑text capability.
- Develop an automated CI/CD agent that interprets build failures, runs appropriate tooling (e.g., lint, tests), and patches code to resolve issues.
- Implement an interactive code reviewer that provides step‑by‑step suggestions and can invoke external tools (e.g., static analysers) to validate changes.
- Design a multi‑modal programming tutor that accepts natural language questions and visual inputs (like error screenshots) to explain concepts and suggest fixes.