Outlier-Ai/Outlier-10B-V2

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Apr 7, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Outlier-Ai/Outlier-10B-V2 is a 30.4 billion parameter (13.3B active) ternary Mixture-of-Experts (MoE) model built upon a frozen Qwen2.5-7B-Instruct base. This version is superseded and retained for reproducibility of earlier benchmark runs, with current research and development focused on its successor, Outlier-10B (V3.3). It represents an archival checkpoint in the development of Outlier-Ai's MoE architectures.

Loading preview...

Outlier-10B-V2: An Archival MoE Checkpoint

Outlier-10B-V2 is a superseded model from Outlier-Ai, primarily maintained for the reproducibility of historical benchmark runs and research. It is an earlier iteration of their Mixture-of-Experts (MoE) architecture, built as a ternary MoE overlay on a frozen Qwen2.5-7B-Instruct base.

Key Characteristics

  • Architecture: Ternary Mixture-of-Experts (MoE) overlay.
  • Base Model: Utilizes a frozen Qwen2.5-7B-Instruct as its foundation.
  • Scale: Features a total of 30.4 billion parameters, with 13.3 billion parameters active during inference.
  • Status: Classified as "archival"; it is explicitly recommended not to use this version for new development or research.

Evolution and Successor

This V2 architecture has been retired in favor of subsequent developments. The successor, Outlier-Ai/Outlier-10B (V3.3), introduced significant architectural changes, including per-expert per-channel scales and ternary TQ1_0 packing. A V3.3 alpha-fix overlay further improved MMLU performance by +1.61 percentage points with a minimal 15 KB addition.

Purpose of Retention

Despite being superseded, Outlier-10B-V2 remains publicly available to adhere to ML research norms, ensuring that external benchmarks and academic papers citing this specific URL can maintain reproducibility. It serves as a historical record of the model's development trajectory.