Outperforms doing what? Inference is not a homogeneous workload, memory bandwidth correlates to decode speed and layer swapping but not necessarily inference speed overall.
The other half of that equation is latency, predicated on prefill performance which needs a powerful GPU and ideally ALU-level optimization to build larger KV caches quickly. Even the M5 gets smoked in this department, the M5 Max has a 50% longer TTFT on Qwen's 27b dense model at only 16k of context, which is a pretty typical starting context to use for agentic editing in normal apps like OpenCode/Claude Code: https://raw.githubusercontent.com/Osmantic/MMBT-Messy-Model-...
For agentic, 50-256k token on-device coding sessions, the Spark will be faster and consume less power running larger models. Without an external GPU (which Apple doesn't support), Apple Silicon will always be bottlenecked during prefill. Apple's failure to address this with their GPU architecture is a big reason why Apple Silicon viewed as a waste of time and money for professional datacenter deployment.
For agentic work, you just cache the prefill kv cache of the relevant system prompts. TTFT is a little slow (10s of seconds, oh no!) the first time you boot up a new harness.
I keep hearing people make this claim that TTFT is a problem, and… it just isn’t, if you’re running oMLX.
Folks in my camp keep saying this, and folks in your camp keep beating a drum we tell you isn’t resonating. Not sure why I keep bothering to argue; you can’t buy a high RAM Mac Studio like mine anymore.
They're not usable for deployment. They're perfectly fine for "enthusiast" low-end usage with 10-30B models, but the same goes for almost every dGPU made in the last 10 years. Your Mac Studio cannot run frontier LLMs at an interactive speed, even Apple has given up on using it as an inference backbone.
