---
format: "aidr-story-markdown/v1"
id: "f9fd05726bee41f9afaaab13464aa03513886e41a8b36adc0775f022d38ee28c"
canonical_url: "https://aidr.today/f9fd0572?lang=en"
title: "Unsloth Shrinks MiniMax's 428B M3.1 Flash Model to Run Locally on 150GB RAM"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-29T19:35:56.000Z"
category: "Models"
topics: ["llm"]
source_urls: ["https://marketbrief.now/ai/unsloth-shrinks-minimaxs-428b-m31-flash-model-to-run-locally-on-150gb-ra-71d914ba","https://x.com/UnslothAI/status/2114730104228286936","https://x.com/AdamHoltererer/status/2115878445045440540","https://x.com/MiniMax_AI/status/2113436622889029887","https://x.com/ValsAI/status/2112403303049117807","https://x.com/teortaxesTex/status/2113447334021644595","https://x.com/ValsAI/status/2112403307075698859","https://x.com/andrew_n_carr/status/2113463883252064567"]
summary: "Unsloth released 2-bit dynamic GGUF quantizations of MiniMax's 428B parameter M3.1 Flash model that reduce its footprint by 45% to 143.9GB and allow it to run locally on 150GB of RAM, retaining 91.5% top 1 token accuracy across seven domains. Independent evaluations showed lower coding benchmark results than MiniMax reported, with two outside tests recording a score of 26 on Terminal-Bench 4.0 compared to MiniMax's claimed 38. The local quantizations follow MiniMax's October 23 release of open weights for M3.1 Flash, a mixture-of-experts model that activates 23B parameters per token."
---

# Unsloth Shrinks MiniMax's 428B M3\.1 Flash Model to Run Locally on 150GB RAM

> [Open the canonical story](<https://aidr.today/f9fd0572?lang=en>)

**Published:** 2026-10-29T19:35:56.000Z
**Category:** Models
**Topics:** llm

## Summary

Unsloth released 2\-bit dynamic GGUF quantizations of MiniMax's 428B parameter M3\.1 Flash model that reduce its footprint by 45% to 143\.9GB and allow it to run locally on 150GB of RAM, retaining 91\.5% top 1 token accuracy across seven domains\. Independent evaluations showed lower coding benchmark results than MiniMax reported, with two outside tests recording a score of 26 on Terminal\-Bench 4\.0 compared to MiniMax's claimed 38\. The local quantizations follow MiniMax's October 23 release of open weights for M3\.1 Flash, a mixture\-of\-experts model that activates 23B parameters per token\.

## Sources

- [Story source](<https://marketbrief.now/ai/unsloth-shrinks-minimaxs-428b-m31-flash-model-to-run-locally-on-150gb-ra-71d914ba>)
- [Story source](<https://x.com/UnslothAI/status/2114730104228286936>)
- [Story source](<https://x.com/AdamHoltererer/status/2115878445045440540>)
- [Story source](<https://x.com/MiniMax_AI/status/2113436622889029887>)
- [Supporting source](<https://x.com/ValsAI/status/2112403303049117807>)
- [Supporting source](<https://x.com/teortaxesTex/status/2113447334021644595>)
- [Supporting source](<https://x.com/ValsAI/status/2112403307075698859>)
- [Supporting source](<https://x.com/andrew_n_carr/status/2113463883252064567>)