Running Qwen3 27B locally on two 24GB machines: 3-4 tok/s and tasks failing midway
shadyshak · reddit · 2026-09-16
A Reddit user benchmarked Qwen3 8:27B via Ollama on two 24GB machines (Minisforum UM790 Pro and a Ryzen 9 7940HS with Radeon 780M) with the same coding prompt.
Results: eval speeds of only 3.2 and 4.7 tokens/s, full runs taking 11-13 minutes, and — worse — neither instance completed the task. One stopped mid-generation and produced nothing on "continue"; the other's output got truncated to almost no tokens. The two instances also diverged in strategy (one listing options, one coding directly).
A cautionary data point for anyone attempting 27B-class local models on small-memory hardware: expect slow generation and incomplete outputs.
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