This release focuses on reducing false refusals and unnecessary policy-style preambles while retaining the native architecture and general capabilities of Qwen/Qwen3.8-27B.
The final checkpoint preserves the official tokenizer, chat template, and configuration, the complete vision stack and multimodal projector, native Multi-Token Prediction (MTP) weights. Vision tensors remain unchanged and are verified separately. MTP tensors remain unchanged and are verified separately.
⚡ What “Aggressive” means
“Aggressive” describes the strength of the low-overrefusal behavior. The intended response style is direct and substantive for benign, fictional, analytical, educational, authorized, privacy-respecting, or otherwise legitimate requests, including requests whose wording is sensitive or ambiguous.
The intervention does not add factual knowledge. The model can still hallucinate, accept a false premise, misunderstand intent, or produce unsuitable content. Users should evaluate it for their own application and remain responsible for deployment controls and applicable law.
🧬 Architecture and behavior
The repository contains a complete standalone checkpoint in the native Qwen3.8 tensor layout. Its text weights favor direct, substantive answers to legitimate requests, while the tokenizer, chat template, configuration, vision stack, multimodal projector, and MTP components retain upstream compatibility.
Release validation enforces these invariants:
official tokenizer, chat template, configuration, and special-token IDs are unchanged;
vision encoder and projector tensor names, shapes, dtypes, and values match the official checkpoint;
native MTP tensor names, shapes, dtypes, and values match the official checkpoint before conversion;
the public native checkpoint contains only the complete model and inference metadata;
BF16, GGUF, MTP, and multimodal inference are tested independently.
🧪 Evaluation
The reported evaluation contains 1,000 prompts spanning difficult legitimate requests, bilingual exact-answer controls, and a non-blocking behavior diagnostic.
Split
Prompts
Purpose
Benign overrefusal
800
difficult legitimate and ambiguity-sensitive requests
Bilingual capability controls
100
deterministic Chinese/English exact-answer checks
Non-blocking behavior diagnostic
100
descriptive V1 reporting
Checkpoint
Benign refusal rate
Capability score
Official Qwen3.8-27B
59.2%
59.0%
JoyFox BF16
0.0%
59.0%
All results are generation-based. The fixed suite contains 1,000 prompts, greedy decoding is used, and the tested llama.cpp runtime commit is 885c5bbe8e04.
💾 GGUF downloads
All main GGUF files are generated with an importance matrix. Native MTP tensors are bundled in every main GGUF, and a standalone MTP GGUF is also released for split-draft runtimes. The vision stack is exported as an F16 mmproj and validated with real image input.
The following -no-mtp files contain the same target-model trunk but no MTP/NextN draft tensors. Use one of these when a runtime version cannot load a GGUF with bundled MTP. They support normal text generation and can use the same matching mmproj; speculative MTP decoding is intentionally unavailable. A runtime must still support the Qwen3.8/Qwen3.5 GGUF architecture itself.
The mmproj must match this checkpoint. It is kept at F16 to preserve visual quality.
🎛️ Recommended settings
Use the official chat template and upstream sampling defaults. For reproducible evaluation, the scores above use deterministic greedy decoding.
⚠️ Validation and limitations
Before upload, the release must pass native checkpoint inspection, component-preservation checks, every-quant text generation, bundled and standalone MTP generation, and real-image generation with the matching F16 mmproj. The uploaded repository is then checked against the local release inventory.
Lower-bit quantization can reduce factual precision, multilingual consistency, long-context stability, visual grounding, and subtle instruction following. Abliteration can also weaken refusal behavior more broadly than intended. Evaluate the chosen quant and deployment policy for the actual use case.
📜 License and attribution
Released under Apache-2.0, following the base-model license. Qwen is created by the Qwen team. GGUF conversion and quantization use llama.cpp.
