Why this matters
Uncensored checkpoints remove the model-internal refusal direction that some alignment pipelines impose; that change can restore latent reasoning capacity but also removes internal safety gates. This HF release packages a lossless BF16 fork of Qwen3.6-27B where the original MTP head from the base model was restored, plus upstream SSM repairs and an abliteration pass tuned to preserve capability while eliminating refusals — a trade that matters for red-team research, security testing, and scenarios where you need the model to follow instructions without alignment-driven refusals.
What Sets It Apart
- Unquantized BF16 reference weights with MTP head grafted from the official Qwen3.6-27B base, preserving the original speculative-decoding head for multi-token-prediction workflows (no retraining required). This makes the BF16 checkpoint the highest-fidelity shipping variant for A100/H100 targets.
- Explicit engineering pipeline: upstream SSM conv1d outlier repair followed by a projected/orthogonal abliteration pass (Optuna-driven) that targets refusal directions while keeping KL divergence to the base model very low. The release documents measured acceptance and speculative-decoding behaviour on production hardware.
- Multiple hardware-aware variants are published alongside (NVFP4 / modelopt formats and XS splits) so operators can choose by memory architecture (dedicated VRAM vs unified memory) rather than by GPU model alone.
Who It's For and Tradeoffs
Great fit if you are running security research, alignment/red-team work, or controlled benchmarking and need a model that reliably executes prompts without internal refusal behavior. Also appropriate for fine-tuning research where a reference BF16 checkpoint is required.
Look elsewhere if you require default safety guardrails, refuse-prone content filtering, or if you cannot accept the legal and operational responsibility of running an uncensored model. This release explicitly removes refusal behavior; the operator bears full responsibility for prompt issuance, downstream use, and compliance with law and policy.
Deployment & Practical Notes
- Use the BF16 variant for A100/H100 (80 GB) serving; NVFP4 / modelopt variants exist for DGX Spark and Blackwell-class cards where FP4 tensor-core throughput is available. Pick variants by memory architecture (unified vs dedicated VRAM) for measured throughput gains.
- The package includes recommended serving settings for vLLM and guidance on concurrency/context window trade-offs; the BF16 checkpoint is ~51–52 GB on disk and intended as the full-precision reference.
- The release maintains provenance and credits the base Qwen model; license is Apache-2.0 inherited from the base.
Risk & Operational Responsibility
This model will generate content the base model would refuse. That includes detailed instructions for harmful activities, graphic content, and other material that many jurisdictions or platforms restrict. Users must implement downstream safety layers (access controls, filtering, audit logs, human review) and accept sole responsibility for prompt design, deployment, and legal compliance. The upstream model card and included clause outline indemnity and arbitration expectations; treat this release as a tool that offloads the safety decision to the operator.
