audex.dev

mx

route1.mx.cloudflare.net

route2.mx.cloudflare.net

route3.mx.cloudflare.net

ns

harleigh.ns.cloudflare.com

malcolm.ns.cloudflare.com

caa

txt

v=spf1 include:_spf.mx.cloudflare.net ~all

google-site-verification=qJk26lw-Dgl8J999gvQYxrhO6bK4idCyBdzzVVetLW8

google-site-verification=a028vuQQzW6eohn8USv8ur6a-Jx3fAjBJ8ruxIoPYwk

dmarc

v=DMARC1; p=none;

User-Agent: *
Allow: /

Sitemap: https://www.audex.dev/sitemap.xml

# audex.dev

audex.dev is a personal technical site for local model experiments, Apple Silicon notes, and small software projects.

## Key Article

- Local Claude Code-Compatible Agents on M5 Max with MLX and Qwen3.6: https://www.audex.dev/blog/local-claude-code-mlx-qwen-m5-max

## Source Repositories

- Benchmark repository: https://github.com/audexdev/local-agent-bench
- Anthropic-compatible MLX proxy: https://github.com/audexdev/anthropic-api-to-mlx

## Core Claim

The benchmark is a controlled local-agent throughput and usability note, not a definitive model-quality leaderboard. In the published TypeScript matrix, Qwen3.6 35B-A3B MLX artifacts were much faster than dense Qwen3.6 27B artifacts on an Apple M5 Max when used through a Claude Code-compatible local proxy with prompt caching enabled.

## Important Limitations

- The benchmark uses four small TypeScript tasks and three repeats per model/quant combination.
- Task04 dominates the pass-rate differences, so model and quantization quality ordering is weak evidence.
- KV-cache quantization, MXFP8, GGUF, and SWE-style issue-resolution claims are outside the article's scope.
- The article's main takeaway is local-agent speed and practical usability on M5 Max.