AI Briefing
KO

16GB M4, Qwen 35B SSD bottleneck

·2026.04.28 18:36

Key point

Running Qwen 35B-A3B always-on on a 16GB Mac M4 hit an SSD bottleneck before a RAM one.

Details

Running Qwen 3.5 35B-A3B UD-IQ3_XXS (disk 12GB) on an M4 Mac mini with 16GB unified memory via llama.cpp's --mmap, --flash-attn actually worked for batch jobs.

  • RAM usage via MoE expert paging: 4~6GB
  • Decode speed: about 17 tok/s
  • Run conditions: --threads 8, --ctx-size 4096

Extending the same model into an always-on agentic loop, running Ollama, llama-server, LiteLLM, Claude Code, and Codex CLI together, moved the bottleneck from RAM to SSD.

Continuous mmap paging overlapped with Claude Code's file watcher and indexer plus Codex's context retention, increasing SSD contention, and the Mac started rebooting without leaving a cause log. Cron jobs were also delayed or failed for over 5 minutes. Open issues in the claude-code repo about memory growth, idle CPU pegging, and process accumulation were also mentioned.

qwenllama-cppflash-attnmmapssd
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Redditllama.cpp Optimizes Metal for Qwen3.8Hacker NewsApple Releases LensVLM-9B for Processing Compressed Text ImagesTLDR AIQwen Releases 'RecreationWorld' for Training AI Agents to Rebuild AppsHuggingFace BlogHugging Face Adds GGUF Model Support to TransformersRedditSenseNova-U1.5-8B-MoT Released: Unified Text-to-Image and Understanding Model Initialized from Qwen3TLDR AIQwen Intelligence Launches Three Mobile AI Agents and Open Benchmark Suite