redashes/Qwen3.8-27B-BF16-SSMFIX-apostate

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 23, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The redashes/Qwen3.8-27B-BF16-SSMFIX-apostate is a 27 billion parameter Qwen3.8 variant, built on an SSMFIX BF16 base, that has undergone selective unlearning using the apostate KCRN method. This model is specifically designed to remove most refusal behaviors while largely preserving its original language model capabilities. It is optimized for applications requiring reduced refusal rates, particularly in sensitive content areas, with minimal impact on general performance.

Loading preview...

Model Overview

redashes/Qwen3.8-27B-BF16-SSMFIX-apostate is a 27 billion parameter model derived from Qwen3.8-27B, specifically the redashes/Qwen3.8-27B-BF16-SSMFIX base. It incorporates a selective unlearning technique called apostate KCRN (Kernel Constraint Rank-Nullity projection) to significantly reduce refusal behaviors.

Key Capabilities & Features

  • Reduced Refusal: Achieves a low overall refusal rate of 7% on harmful prompts, with remaining refusals concentrated in legally sensitive areas like PII/privacy and self-harm.
  • Capability Retention: Benchmarks show strong retention of knowledge (CMMLU), math reasoning (GSM8K), and even improved truthfulness in generation (TruthfulQA gen) compared to the official BF16 base.
  • Analytic Unlearning: Utilizes a closed-form, analytic unlearning algorithm that structurally guarantees refusal information removal while preserving normal generation directions, avoiding iterative tuning.
  • High Context Length: Supports a context length of 262144 tokens, inherited from its base.
  • BF16 Precision: Operates in BF16 (bfloat16) precision, offering a balance of performance and memory efficiency.

Why Use This Model?

This model is particularly suitable for use cases where:

  • Minimizing Refusal is Critical: Applications that require a language model to be less prone to refusing prompts, especially in non-harmful but sensitive contexts.
  • Maintaining Core Capabilities: Users need a model that retains strong general language understanding, reasoning, and generation abilities despite unlearning specific behaviors.
  • Experimental Research: Developers interested in exploring advanced unlearning techniques and their practical impact on LLM behavior.

Finetunes
1 models