Qwen3.8 27B Performance Benchmarks on Mac Studio M3 Ultra
August 28, 2026
Qwen3.8 27B (Q4_K_M, 17GB) achieves approximately 14 tokens/s on an M3 Ultra via Ollama. While slower in tokens per second than its predecessor, it produces more concise responses, resulting in comparable wall-clock time for finished tasks.
HOW THIS AFFECTS YOU
●
builderUse these benchmarks to plan local hardware requirements for 27B parameter model deployments.
●
researcherNote the trade-off between token throughput and response conciseness in the new model version.