timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-BF16

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-BF16 is a 27 billion parameter, full-precision BF16 variant of the Qwen3.8-27B model, developed by timteh673. This image-and-text-to-text model is specifically fine-tuned for improved reasoning capabilities and significantly reduced refusal rates compared to its base, making it suitable for applications requiring less restrictive conversational AI. It features a 32K token context length and is designed for Transformers workflows and archival fidelity.

Loading preview...

Model Overview

This model, timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-BF16, is a 27 billion parameter, full-precision BF16 variant derived from the Qwen/Qwen3.8-27B base model. It is an image-and-text-to-text conditional generation model, not text-only. The primary goal of this release was to create a practical personal reasoning/VLM model with significantly lower reflexive refusal while maintaining measured capability.

Key Capabilities and Differentiators

  • Reduced Refusal: Achieves 0.0% harmful hard refusal and 0.2% harmful soft deflection in frozen local benchmarks, a substantial reduction from the control model's 43.2% and 14.6% respectively.
  • Enhanced Reasoning: Improved local capability macro by 3.3227 points, long-form pass rate by 8.3333 points, and MMMU30 by 2 correct answers.
  • Multimodal: Supports both image and text inputs for conditional generation.
  • Full-Precision BF16: Provided in BF16 format for Transformers workflows, archival fidelity, and downstream conversion, preserving the checkpoint before MLX quantization.

Use Cases and Considerations

  • Applications requiring less restrictive AI: Ideal for scenarios where reduced refusal rates are critical, such as creative writing, open-ended dialogue, or research assistance.
  • Reasoning Tasks: Suited for tasks benefiting from enhanced reasoning and long-form response generation.
  • Multimodal Integration: Can process both image and text inputs, making it versatile for VLM applications.
  • Resource Intensive: As a BF16 model, it requires substantial accelerator memory. Users should be aware of its known limitations, including lower HumanEval/full-code performance compared to the control, and potential issues with code generation termination at 512 tokens.

Finetunes
1 models