---
title: "Discussion on \"Local models are actually good now - playing with Qwen3.8-27B\" | Hashnode"
description: "Discussion on \"Local models are actually good now - playing with Qwen3.8-27B\". A matched Pi experiment shows why quantisation is an agent memory decision: 4-bit and 6-bit finished, while 8-bit hit Metal OOM."
url: "https://hashnode.com/posts/qwen3-8-27b-on-m4-pro/6aa1c2f1e9d978acd85efffb"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Local models are actually good now - playing with Qwen3.8-27B](https://billiem.hashnode.dev/qwen3-8-27b-on-m4-pro?utm_source=hashnode&utm_medium=feed)

A matched Pi experiment shows why quantisation is an agent memory decision: 4-bit and 6-bit finished, while 8-bit hit Metal OOM.

Sep 9 · 7 min read · billiem.hashnode.dev  
Read the article

-   [#local-llms](https://hashnode.com/tag/local-llms)
-   [#machine-learning](https://hashnode.com/tag/machine-learning)
-   [#apple-silicon](https://hashnode.com/tag/apple-silicon)
-   [#mlx](https://hashnode.com/tag/mlx)
-   [#ai-agents](https://hashnode.com/tag/ai-agents)

## Written by

[Billie M](https://hashnode.com/@billiem)

@billiem

  
[View profile](https://hashnode.com/@billiem)

## Responses

Markdown supported · ⌘↵ to post

No responses yet.