Picking Models as a Mac User

I recently spent a couple of weeks replacing models around my house, which led me to think that sharing my model selection criteria could be helpful for Mac users. When a new model launches, I start by checking forums and comments to gauge real-world performance. Benchmarks are useful, but seeing how people actually use the model is crucial.

I look for issues with tokenizers, bug-bitten implementations, and any overthinking or hallucinations. I also check Artificial Analysis, which has benchmarks aligned with my needs, such as context reasoning, hallucination rate, and output tokens. On a Mac with limited VRAM, I'm particularly conscious of models generating many tokens, as it can slow things down significantly.

Qwen3.8 27B is a prime example - while it matches Opus 4.6 in benchmark performance, its high token generation needs make it less practical unless you're okay with waiting for long response times. I compared a few smaller models suitable for M2 Ultra and M5 Max Macs: Qwen3.8-27B offers impressive performance but at a high token cost; Muse Glimmer High provides good reasoning and output tokens at 48M for 30B; Gemma 4 31B excels in context and tone, running well with reasoning off; GLM-5.3-Flash has a high intelligence score but fewer active parameters than some others.

For my M3 Ultra 512GB, MiniMax M3 stands out as the top choice despite having fewer active parameters compared to some alternatives. While higher-scoring models like GLM-5.3-Flash or GLM-5.2 Max might seem appealing based solely on intelligence, MiniMax M3's balanced performance across all metrics makes it the best fit for my needs, especially considering my preference for keeping output token usage low.