Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16
Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16 is a 27 billion parameter dense hybrid Vision-Language Model (VLM) developed by Blackfrost, based on the Qwen3.8 architecture. This model is provided in native BF16 precision and retains vision, reasoning, tool-use, and long-context capabilities. Its primary differentiator is a weight-level modification to significantly reduce false-positive refusals, making it suitable for lawful software engineering and security tasks.
Loading preview...
Overview
Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16 is a 27 billion parameter dense hybrid VLM, built by Blackfrost on the official Qwen3.8-27B base. This model is provided as a native BF16 safetensors checkpoint and is a weight-level derivative, not a fine-tune or merge. It supports text, image, and video inputs with text output, and features a substantial context length of 262,144 tokens natively.
Key Capabilities & Differentiators
- Reduced Refusal Surface: The core innovation is a deliberate weight-level modification to reduce false-positive refusals, achieving a residual refusal rate of 2.4% across 450 test cases (AdvBench, StrongREJECT, XSTest). This makes it particularly useful for applications requiring less restrictive responses in lawful contexts.
- Retained General Capabilities: Despite the refusal surface modification, the model aims to retain its original vision, reasoning, tool-use, and long-context capabilities from the Qwen3.8 base.
- Deployment Friendly: Designed as a dense, deployment-friendly model, it is validated for serving on single NVIDIA B200 GPUs for conservative 8K context serving.
- Native BF16 Precision: Ships in native BF16 safetensors format, ensuring high precision.
Should I use this for my use case?
This model is ideal for developers and researchers working on applications where reducing unwarranted refusals is critical, especially in software engineering and security domains. If your use case involves tasks that might trigger false-positive safety refusals in other models, but require lawful and operator-controlled responses, this model's modified refusal surface could be highly beneficial. It is a research preview, and users should be aware that it is not a safety-stock model and requires external controls for production use.