Skip to content
Trending storyDeveloping

Zyphra's ZHA method converts Qwen3.6-35B-A3B with 81,920 trainable parameters

1 article1 sourcesince Nov 2Last article 2h ago ·

Overview

AISummary of 1 article

Zyphra says its ZHA method converts existing pretrained models through distillation using only a few billion tokens while keeping the base model frozen, so no training from scratch is needed.

For Qwen3.6-35B-A3B, Zyphra says ZHA adds only 81,920 trainable parameters, about 0.00024% of the model.

The report is a single post on Zyphra's X account, and it does not give benchmark results for the converted model.

Written by AI from the articles below · updated Nov 2, 5:23 PM ET

Check the sources:

Latest developmentNov 2, 5:04 PM ET
Zyphra converts pretrained models with ZHA using 81,920 trainable parameters

Article timeline

The articles in this story. Times are ET.

Nov 2
  1. ZyphraOfficial
    Zyphra converts pretrained models with ZHA using 81,920 trainable parameters

    AIZyphra says it converts existing pretrained models through distillation using only a few billion tokens while keeping the base model frozen, so no training from scratch is needed. For Qwen3.6-35B-A3B, its ZHA method adds only 81,920 trainable parameters, about 0.00024% of the model.

Heat Trend

Not enough continuous observations to show a trend yet.