Why this matters
This release packages a deployment-oriented, quantized variant of Qwen3.8-27B for local inference workflows. Rather than a fine-tune, the checkpoint applies a weight-level modification to the model's refusal surface and then exposes a complete, single-file GGUF quant ladder (Q2_K through Q8_0) with the model's MTP/NextN drafting head embedded in each main quant. That combination targets users who need an easy-to-load, multimodal 27B model for llama.cpp-style serving while accepting the behavioral tradeoffs of a deliberately reduced-refusal artifact.
Key Capabilities
- One-file GGUF ladder: every standard K-quant is provided so you can choose memory/quality tradeoffs from ~10.9 GB (Q2_K) to ~29.0 GB (Q8_0) without sidecar files.
- Embedded speculative drafting (MTP): the NextN/MTP block is baked into each main quant, enabling speculative/draft decoding without loading separate draft sidecars — improves compatibility with llama.cpp and llama-server workflows.
- Multimodal support via vision projectors: optional mmproj files (F16 and Q8_0 projectors) let you enable image/video input while keeping the text quant separate.
- Reduced-refusal derivative: refusal behavior was modified at the weight level to lower operational refusals (release benchmark reported 11/450 residual refusals under a specified template), but this is not a safety fine-tune and must be validated per workload.
Who it's for and tradeoffs
Great fit if you need a runnable, local multimodal 27B that: loads directly in llama.cpp/llama-server, supports large contexts (architectural 262k tokens), and lets you pick precise quant/size tradeoffs for constrained hardware. The embedded MTP head simplifies speculative decoding setups and reduces chance of sidecar-loading issues.
Look elsewhere if you require an unmodified upstream behavior or guarantees about safety boundaries: the artifact intentionally alters refusal behavior at the weight level and ships an embedded execution prompt in the default Jinja chat template. It also does not include pruning, SFT, LoRA, or secure deployment tooling — operators must provide access control, tool credentialing, and runtime isolation themselves.
Practical notes
- Pick one text quant plus one mmproj for vision input; the repository provides both full-fidelity and compact projectors.
- Validate sampling, context length, and any tool integrations in your target serving stack before production.
- License: remains under Apache-2.0 as inherited from the Qwen3.8 base checkpoint.
