Why this matters
Qwopus3.6-27B-v2 targets a common fine-tuning failure mode: small-to-mid sized student models imitating compressed “reasoning bubbles” produced by closed APIs end up with brittle, shortcut-driven logic. This release uses Trace Inversion to reconstruct step-by-step chains-of-thought and then applies a three-stage curriculum SFT to teach those learnable traces to a 27B dense model — provided as a GGUF artifact with vision & tool-use formatting.
Key Capabilities
- Structured reasoning distillation: incorporates two Trace Inversion datasets (claude-opus-4.6/4.7 inverted sets) to replace compressed summaries with explicit, learnable CoT traces — improves token-efficiency and reasoning continuity.
- Multimodal & long-context support: packaged for image-text-to-text pipelines, MTP head handling, and staged SFT to scale context windows up to 32K tokens.
- Practical engineering: custom Qwen MTP splitting/merging tooling (qwen-mtp-gguf) and Unsloth-accelerated fine-tuning pipeline; distributed/hardware notes and tips included for secondary fine-tuning.
Who it’s for, and trade-offs
Great fit if you want a research-oriented, locally runnable 27B dense model in GGUF that emphasizes explicit chain-of-thought behavior for complex reasoning, multimodal prompts, or agentic workflows. It’s useful for experimenting with Learnable CoT, tool-calling formats, and long-context agentic tasks.
Look elsewhere if you need a production-ready, safety-vetted commercial model: this is a community experimental release (limited safety evaluation, non-standard benchmarking) and may require dependency pinning, LoRA/merge care, and manual patches for large-scale merges or further fine-tuning.
Where it fits
Compared with an unmodified Qwen3.6-27B checkpoint, the release claims improved reasoning accuracy on targeted subsets (example: reported MMLU‑Pro subset uplift) and token-efficiency gains through structured CoT supervision and MTP speed tuning. Use the dense 27B variant for deep, stepwise reasoning and long-context agentic uses; choose MoE/high-throughput options if raw token throughput is the priority.
Implementation notes (concise)
- Format: GGUF (download includes mmproj.gguf for vision).
- Base: qwen/Qwen3.6-27B; fine-tuned with Unsloth and a three-stage curriculum pipeline.
- Known issues: LoRA weight merges can hit OOM; dependency compatibility (PEFT / Transformers / Unsloth) may require specific version pins; community release—exercise caution for deployment.
If you want, I can: (1) extract key file names and download sizes from the model page, (2) summarize the repository’s finetuning scripts (Jackrong-llm-finetuning-guide), or (3) produce example inference prompts to test the model’s CoT behavior.
