AI BriefingQwen3.5-35B Fast Performance
Key point
On an M2 Ultra 64GB, Qwen3.5-35B-A3B Q8_K_XL achieved 63 t/s in generation.
Details
This is a benchmark result from running Qwen3.5-35B-A3B Q8_K_XL on a Mac Studio M2 Ultra 64GB.
- Prefill 10240: 1734 t/s
- Prefill 16384: 1552 t/s
- Generate 512: 63 t/s
The run parameters were -ngl 99 -fa 1 -b 2048 -ub 2048 -ctk bf16 -ctv bf16 -mmp 0, and the results were aggregated as an average of 3 runs.
Even in the long-prompt range, prefill remained in the 1,500-1,700 t/s range, and the generation stage recorded 63 t/s, providing a real-world measurement for gauging local large-model inference performance.
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
Qwen Intelligence Launches Three Mobile AI Agents and Open Benchmark Suite
Apple Releases LensVLM-9B for Processing Compressed Text Images