DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 9, 2026License:apache-2.0Architecture:Transformer0.1K Open Weights Featherless Exclusive Cold

DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored is a 27 billion parameter Qwen 3.8-based model developed by DavidAU, featuring a 32K context length. This model is engineered for superior instruction following and detail, significantly reducing "thinking tokens" while maintaining high intelligence. It offers multiple reasoning and instruct modes switchable on the fly, and is specifically designed as an "ULTRA HERETIC" version with strong de-censoring capabilities, achieving only 6/100 refusals in testing.

Loading preview...

Model Overview: DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored

This 27 billion parameter model, built on the Qwen 3.8 architecture by DavidAU, introduces "TWIN-TURBO" technology to drastically reduce "thinking tokens" (by 1/2 to 1/20) while enhancing performance and intelligence. It features 5 distinct thinking modes and 5 instruct modes, which are switchable dynamically via API or directly within chat messages.

Key Capabilities & Differentiators

  • Superior Instruction Following: Internally named "Stage2b-rplus3," this model was selected for its consistent instruction adherence, attention to detail, and reliable generation quality.
  • Reduced Refusals: As an "ULTRA HERETIC" version, it exhibits strong de-censoring, with a refusal rate of only 6/100, significantly lower than the base Qwen 3.8 (86/100).
  • Enhanced Reasoning: Excels in deep detail, double-checking, and multi-stage drafting when prompted, indicating advanced reasoning capabilities.
  • Benchmark Performance: Demonstrates competitive performance across various benchmarks (ARC-C, ARC-E, BoolQ, HSWAG, OBKQA, PIQA, WINO) compared to base Qwen 3.8 and other Qwen variants, particularly in mxfp8 and mxfp4 quantizations.

Good For

  • Applications requiring highly uncensored and unfiltered responses.
  • Tasks demanding precise instruction following and detailed output.
  • Scenarios where efficient token usage for reasoning is critical.
  • Use cases benefiting from dynamic switching between different reasoning and instruction styles.

Finetunes
5 models