A naukri.com initiative
Macstories
2w
365
Image Credit: Macstories
Notes on Early Mac Studio AI Benchmarks with Qwen3-235B-A22B and Qwen2.5-VL-72B
- The author received a top-of-the-line Mac Studio on loan from Apple to experiment with assistive AI and automation.
- Utilizing the power of Apple Silicon and the M3 Ultra model, the author tested local LLMs to explore their performance.
- Early benchmarks with the Qwen3-235B-A22B model showed better performance on the Apple Silicon-optimized MLX version.
- The author also experimented with vision models like Qwen2.5-VL for OCR tasks, achieving good results locally.
- Cloud models often struggle with image compression, but local VLMs on Mac Studio provided accurate results quickly.
- The author built a local scanning utility to run images through Qwen2.5-VL, showcasing the efficiency of local VLMs.
- The article highlights the potential of hybrid automation using Mac scripting tools, Shortcuts, and LLMs.
- Future plans include exploring MLX models, long-context models, and Gemma 3 with the power of the Mac Studio.
- Club MacStories offers exclusive content, newsletters, a Discord community, app discounts, and more for members.
- Different membership tiers provide access to various perks like early podcast releases and ad-free content.
- Join Club MacStories to explore a range of benefits and engage with the MacStories community.
Read Full Article
21 Likes
For uninterrupted reading, download the app