---
format: "aidr-story-markdown/v1"
id: "a89e75e880101985bfc50faa367b8488eb72e48b4b499c5007f34e7c78745d67"
canonical_url: "https://aidr.today/a89e75e8?lang=en"
title: "I ran 120B MoE models on my 32GB M1 Max using under 16 GB"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-09T19:20:34.000Z"
category: "Infra"
topics: ["inference"]
source_urls: ["https://mitln.github.io/moe-inference-engine/","https://news.ycombinator.com/item?id=50430323"]
summary: "A prototype was created to run gpt-oss-120b and Qwen3.5-122B-A10B on a MacBook Pro with an M1 Max and 32GB of RAM using a 16 GiB process-memory budget. On gpt-oss-120b, the engine averaged 7.84 tokens/s compared to 5.78 tokens/s for TUFF 10.2.2 in the same setup."
---

# I ran 120B MoE models on my 32GB M1 Max using under 16 GB

> [Open the canonical story](<https://aidr.today/a89e75e8?lang=en>)

**Published:** 2026-11-09T19:20:34.000Z
**Category:** Infra
**Topics:** inference

## Summary

A prototype was created to run gpt\-oss\-120b and Qwen3\.5\-122B\-A10B on a MacBook Pro with an M1 Max and 32GB of RAM using a 16 GiB process\-memory budget\. On gpt\-oss\-120b, the engine averaged 7\.84 tokens/s compared to 5\.78 tokens/s for TUFF 10\.2\.2 in the same setup\.

## Sources

- [Story source](<https://mitln.github.io/moe-inference-engine/>)
- [Discussion](<https://news.ycombinator.com/item?id=50430323>)