““Wake me up when they release the training data and pipeline.” — NNN_Throwaway2, Reddit: https://www.reddit.com/r/LocalLLaMA/comments/1v275de/davidau_somehow_managed_to_improve_qwen_36_27b/”
You know that feeling when a local model refuses a request you want it to answer, or when a capable model is too large for your hardware? This project targets both pains by removing most refusals from Qwen3.6-27B and distributing quantized files for local runtimes. You can use it for reasoning, coding, writing, roleplay, and vision without relying on an inference provider. The trade is that memory use, CPU offload, and token speed vary sharply by quant and device.
You start with Qwen3.6-27B, then the team applies multiple fine-tuning stages and merges. Heretic with Arbitrary-Rank Ablation removes much of the base model's refusal behavior, while Fable, Claude Opus reasoning, F451, and GPT5 Polaris material are named as parts of the training mix. The result is packaged as GGUF files with regular and MTP variants; MTP predicts extra tokens and can be faster when acceptance is high. You add a separate `mmproj` file when you want vision input.
If you want a locally hosted model for coding, reasoning, creative writing, roleplay, or image input, this is worth a hardware-specific test. You should skip it as a default production choice if you need predictable speed across AMD and NVIDIA devices, a fully reproducible training pipeline, or a broad independent benchmark.
Worth exploring as an experimental local model, not as a drop-in replacement for a hosted frontier model. The card gives you runnable commands, large download activity, and concrete benchmark data, while community reports flag a small benchmark sample, incomplete training artifacts, and serious AMD offload behavior.
Deep-dive insight, Easy and Pro modes, plus action playbooks — the full breakdown is one tap away.