AnalysisAI ModelsJuly 3, 2026

Qwen 27B achieves 50-90 tokens/s on local hardware

A user reports Qwen 27B at q6kxl with multi-token prediction achieves 50-90 tokens/s decode and 1500-2200 tokens/s pre-fill on a 4090+3090 system using lcpp. The model interfaces reliably with any API when provided with documentation. This demonstrates strong local performance for a 27B model.

1 source

Qwen 27B achieves 50-90 tokens/s on local hardware — AIBriefs