open-machine/Qwen3-8B-FlashNorm

TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 22, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

open-machine/Qwen3-8B-FlashNorm is an 8 billion parameter language model, mathematically equivalent to Qwen/Qwen3-8B, but optimized using FlashNorm for faster inference. Developed by OpenMachine, this model integrates RMSNorm weights directly into linear layers, removing them from the state dictionary. This optimization maintains original model output while enhancing computational efficiency, making it suitable for applications requiring high-speed transformer inference.

Loading preview...

Qwen3-8B-FlashNorm Overview

This model is a FlashNorm-prepared checkpoint of the original Qwen/Qwen3-8B model, developed by OpenMachine. It implements the FlashNorm technique, which is an exact reformulation of the RMSNorm -> Linear operation, as detailed in the paper FlashNorm: Fast Normalization for Transformers.

Key Capabilities & Optimizations

  • Mathematical Equivalence: Despite modifications, the model produces the same output as the original Qwen3-8B, as proven by Proposition 1 of the FlashNorm paper.
  • Performance Enhancement: FlashNorm folds per-channel RMSNorm weights into subsequent linear layers, removing them from the state dictionary. This pre-computation aims to accelerate inference by simplifying the normalization step at runtime.
  • Reduced State Size: The removal of norm weight tensors from the state dict can lead to a more streamlined model representation.

Usage & Compatibility

  • HuggingFace Transformers: The model is fully compatible with HuggingFace Transformers, though it will issue a warning about missing norm weights, which is expected and correct behavior for FlashNorm.
  • vLLM Support: Currently, vLLM does not natively support this checkpoint due to the absence of norm weight tensors, but a patch is being tracked.
  • Local Regeneration: Users can regenerate a FlashNorm-prepared checkpoint locally using the transformer_tricks library.

Good For

  • Developers seeking faster inference for Qwen3-8B without compromising output quality.
  • Applications where computational efficiency and reduced model footprint are critical.
  • Experimentation with advanced transformer optimization techniques like FlashNorm.

Finetunes
1 models