DavidAU/ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2

TEXT GENERATIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:21BQuant:FP8Context Size:32kPublished:Feb 19, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

DavidAU/ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2 is a 21 billion parameter MOE (Mixture of Experts) model, specifically an uncensored fine-tune of the Ernie 21B-A3B architecture with 64 experts. Fine-tuned using a Gemini 3 Pro High Reasoning dataset, this model excels in creative tasks like brainstorming and prose generation, as well as general usage. It features a 128k context window and provides compact yet detailed reasoning, with stable reasoning across a temperature range of 0.1 to 2.5.

Loading preview...

Model Overview

DavidAU/ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2 is a 21 billion parameter Mixture of Experts (MoE) model based on the Ernie 21B-A3B architecture, featuring 64 experts. This model is an uncensored fine-tune, developed using a Gemini 3 Pro High Reasoning dataset, and is designed for deep thinking capabilities.

Key Capabilities & Features

  • Enhanced Reasoning: Provides compact yet highly detailed reasoning, which is stable across a temperature range of 0.1 to 2.5.
  • Creative & General Use: Excels in creative tasks such as brainstorming and generating creative prose, alongside strong performance in general usage scenarios.
  • Uncensored Output: Noted for being largely uncensored "right from the factory."
  • Extended Context Window: Supports a 128k token context length.

Performance & Benchmarks

The fine-tuning process has resulted in improved performance across several benchmarks compared to the regular model:

  • arc_challenge: 0.373 (vs 0.331)
  • arc_easy: 0.444 (vs 0.440)
  • hellaswag: 0.679 (vs 0.663)
  • openbookqa: 0.362 (vs 0.338)
  • piqa: 0.758 (vs 0.725)
  • winogrande: 0.642 (vs 0.567)

Optimal Usage Settings

For best results, especially in creative or chat/roleplay scenarios, specific settings are recommended:

  • Repetition Penalty: Suggests 1.01 to 1.1 for creative tasks, and 1 (off), 1.05, or 1.1 for general work.
  • Smoothing Factor: Setting Smoothing_factor to 1.5 in interfaces like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern is advised for smoother operation and to mitigate looping issues with lower quantizations. This is also referred to as "Quadratic Sampling."

For comprehensive guidance on advanced parameters and samplers to maximize model performance, users are directed to the Maximizing Model Performance guide.

Finetunes
1 models