Framework Desktop with 128GB unified memory: Qwen 27B-35B beats bigger local models
lebbi · reddit · 2026-09-17
The author shares hands-on experience running local LLMs on a Framework Desktop with AMD Strix Halo and 128GB of unified memory:
- Counterintuitive finding: despite 90GB of spare memory, mid-sized models like Qwen 3.6 35B and Qwen 3.8 27B consistently give the best coding results and end up being the daily drivers;
- The bottleneck is CPU: it can't run multiple agents concurrently at decent speed, so much of that memory sits idle;
- The author asks the community what they actually run on Strix Halo machines and whether they're missing something.
Useful reference for anyone considering big-memory machines for local LLMs: more memory doesn't mean bigger models — speed and CPU often limit you first.
More from Infra
- Palmer Luckey asks: can you use datacenter RAM for things prohibited by the operator's terms of service? — avaitopiper · 2026-11-02
- Disable MTP for 20-30% faster local LLM inference on limited RAM — InternalMode8159 · 2026-11-02
- Kioxia up 216% in a year but down 68% from its June 2026 peak — zephyr_z9 · 2026-11-02
- Intel plunges 29%, worst day since the 1970s, wiping $12B off US government stake — emmanuelvivier · 2026-11-02
- Researcher hoards nearly 1TB of DRAM, refuses to sell a stick to data centers — kevinnbass · 2026-11-02
- Self-Hosting LLMs at a Law Firm: Infra and Data Governance Harder Than the Model — EducationalGroup7245 · 2026-11-02