@karminski3: Thinking of buying a Mac to run large models? This is a deterrent post. Actually, the estimation method is simple. Even if you buy a MacStudio to run the Qwen3.6-27B 4bit quantized version, then enable DFlash to use Qwen's built-in speculative decoding, it only reaches 65 token/s. And now most large models can run at 40 token/s…

Thinking of buying a Mac to run large models? This is a deterrent post. Actually, the estimation method is simple. Even if you buy a MacStudio to run the Qwen3.6-27B 4bit quantized version, then enable DFlash to use Qwen's built-in speculative decoding, it only reaches 65 tokens/s. And now most large models can run at 40 tokens/s. If you specifically buy a MacStudio M3 Ultra 96G to run large models, and convert the device price (32,999) into API costs, taking GLM-5.2 as an example, at 28 RMB per million tokens, the price of one MacStudio can buy about 32,999 / 28 = 1,178 million tokens. To output these tokens, the MacStudio running Qwen3.6-27B would need to run continuously for 209 days. That means the break-even period is at least 200 days of uninterrupted operation. Only after that, running models is pure profit. This doesn't even include electricity costs or the fact that you could buy API packages instead of pay-as-you-go. And most importantly, this is only for a small 27B model. If you actually buy a 512G MacStudio (108,749, and it seems to be out of stock), and run a quantized version of GLM-5.2, the speed drops to only 17 tokens/s, and the break-even period is about 7 years... Given that new model versions are released every 1.5 months, it's definitely not cost-effective for ordinary users to buy for personal use. So for most users, buying a coding plan is more cost-effective. If, like me, you need to test new models, renting GPUs directly is much more economical than buying. Of course, if you already own a Mac or GPU, then letting it run large model tasks during idle time (e.g., while sleeping) is actually cost-effective. #LocalLargeModels #mac #qwen36 #glm52