1.6 trillion-parameter DeepSeek model demoed running on a MacBook M5 Max via SSD

A demonstration showed a 1.6 trillion-parameter DeepSeek model — reportedly a variant of DeepSeek v4 PRO — running on a MacBook M5 Max by streaming weights from SSD. The feat is striking because a model of that scale would normally require data-center GPUs, not a laptop.

The technical trick is memory hierarchy: rather than holding all 1.6T parameters in RAM, the setup pages weights from fast SSD storage, trading some throughput for the ability to run an enormous model on consumer hardware. It's a proof-of-concept for how far on-device inference can be pushed with sparse/MoE architectures and clever memory management.

The story fits squarely into the week's dominant local-AI theme — Google's Gemma 4 on a 16GB laptop, AI Edge Gallery on macOS, NVIDIA RTX Spark, and Intel/Perplexity's on-device pushes. The appeal is privacy, offline capability and zero per-token cost; the trade-off is speed.

The heavy caveat: SSD-streaming a 1.6T model will be slow, so this is more a demonstration of possibility than a practical daily-driver setup. Sourcing is also thin (aggregator posts), so independent confirmation of throughput and exact configuration is warranted before reading too much into it. Still, it underscores the direction of travel: frontier-scale capability creeping toward the edge.