Qwen 3.8 27B: LM Studio vs MLX on an M3 Max

This August 2026 Spanish-language comparison tests Qwen 3.8 27B on a MacBook Pro M3 Max with 48 GB of unified memory. The presenter uses LM Studio to compare GGUF quantizations, then tests an MLX version through Ollama. The focus is local LLM inference speed and the memory available for context. The presenter installs asitop with pip and starts it with sudo to inspect Apple Silicon CPU and GPU cores, memory and swap. He returns to asitop during inference to compare GPU utilization, memory use and power consumption.

In LM Studio, the presenter loads all 65 layers and sets Q8 to a 16,000-token context. A short French Revolution summary returns 18.53 tokens per second; a longer astronaut story falls to 12.10. He suspects memory pressure and swap contribute to the slowdown, but does not establish the cause. Q6 allows roughly 64,000 tokens of configured context and returns 10.96 tokens per second on the story prompt. Q4 uses a 128,000-token setting and reaches 19.14 on that prompt.

The MLX test reports 39.76 tokens per second for the summary and 42.59 for the story. The presenter checks a configured context of about 262,000 tokens. These short generation tests do not demonstrate performance with that entire window filled.

An RTX 5080 with 16 GB of video memory produces an unresolved result. The presenter says the Q4 file exceeds 17 GB, so he tries Q3 variants with a 10,000-token context. He reports no more than 20 tokens per second and GPU utilization around 30%. A runtime update does not resolve the issue; he leaves diagnosis for a later video.