---
format: "aidr-story-markdown/v1"
id: "572de4e5a0081c0eb88f134dd0c4b4b84ec6700b78cf68c536cebbf2ee7943fd"
canonical_url: "https://aidr.today/572de4e5?lang=en"
title: "Prime Intellect's New MoE Kernel Speeds Qwen3 Model Training by Up to 55%"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-29T17:52:27.000Z"
category: "Frameworks"
topics: ["deepseek","llm","framework"]
source_urls: ["https://marketbrief.now/ai/prime-intellects-new-moe-kernel-speeds-qwen3-model-training-by-up-to-55p-9ec3e997","https://huggingnews.com/ai/prime-intellects-new-moe-kernel-speeds-qwen3-model-training-by-up-to-55p-9ec3e997"]
summary: "Prime Intellect released Prime Mega MoE, a tool that optimizes training of large-scale Mixture-of-Experts AI models, reportedly accelerating Qwen3-235B-A22B training by 55% and Qwen3-30B-A3B by 48% compared to TorchTitan."
---

# Prime Intellect's New MoE Kernel Speeds Qwen3 Model Training by Up to 55%

> [Open the canonical story](<https://aidr.today/572de4e5?lang=en>)

**Published:** 2026-10-29T17:52:27.000Z
**Category:** Frameworks
**Topics:** deepseek, llm, framework

## Summary

Prime Intellect released Prime Mega MoE, a tool that optimizes training of large\-scale Mixture\-of\-Experts AI models, reportedly accelerating Qwen3\-235B\-A22B training by 55% and Qwen3\-30B\-A3B by 48% compared to TorchTitan\.

## Sources

- [Story source](<https://marketbrief.now/ai/prime-intellects-new-moe-kernel-speeds-qwen3-model-training-by-up-to-55p-9ec3e997>)
- [Story source](<https://huggingnews.com/ai/prime-intellects-new-moe-kernel-speeds-qwen3-model-training-by-up-to-55p-9ec3e997>)