---
format: "aidr-story-markdown/v1"
id: "cfa1a627bd09ffd3bddab6820b2921cedfd294a922a81118c0634306d9e7ccce"
canonical_url: "https://aidr.today/cfa1a627?lang=en"
title: "Last year we built megakernels. This year we deleted them: The road to Together Compiler"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-16T00:00:00.000Z"
category: "Frameworks"
topics: ["inference","framework"]
source_urls: ["https://www.together.ai/blog/road-to-together-compiler"]
summary: "Together Compiler uses AI agents to rewrite and fuse GPU kernels in a model's forward pass. On a B200, it reportedly beats hand-written megakernels by 28%, achieving 2,118 tokens/s compared to 1,649 tokens/s on Llama 3.2 1B."
---

# Last year we built megakernels\. This year we deleted them: The road to Together Compiler

> [Open the canonical story](<https://aidr.today/cfa1a627?lang=en>)

**Published:** 2026-11-16T00:00:00.000Z
**Category:** Frameworks
**Topics:** inference, framework

## Summary

Together Compiler uses AI agents to rewrite and fuse GPU kernels in a model's forward pass\. On a B200, it reportedly beats hand\-written megakernels by 28%, achieving 2,118 tokens/s compared to 1,649 tokens/s on Llama 3\.2 1B\.

## Sources

- [Story source](<https://www.together.ai/blog/road-to-together-compiler>)

