Two Years of Local AI on a Laptop

Two years ago, the smartest open-weight model that fit on a 128 GB MacBook Pro was Llama 3 70B. Its score on the Artificial Analysis Intelligence Index was 10.

Today, on the same MacBook Pro, the smartest one that fits is DeepSeek V4 Flash, running in custom 2-bit quantization (about 81 GB of weights). Score: 47.

That is 2.23 doublings in 24 months on a laptop that has not changed. The hardware is barely the story. Every gain is in model design, training data, and post-training.

Three mechanisms compounded over those 24 months. Mixture-of-Experts architectures (DeepSeek V3, V4, Qwen3-Next) keep total parameter count high while only activating a fraction per token, so inference stays fast at scale while the model gets smarter. Four-bit quantization stopped being a quality sacrifice and became the default for most local serving, quartering memory versus the original bf16 weights with small accuracy cost on most tasks. Distillation from frontier teachers became routine, so a small open-weight model now inherits behavior that took a much larger teacher to develop.

What that means today on a 128 GB MacBook: an agentic coding assistant for daily work, long-context retrieval over your private docs, a code reviewer that matches GPT-4 class from late 2024. The infrastructure question moved from "can I run this" to "which one fits my task".

The next 24 months will compound similar gains. Architecture (MoE, Multi-Token Prediction, Multi-head Latent Attention) and post-training methods evolve faster than silicon. Each is reducible to engineering effort, not fabrication cycles. The gap that today still requires the cloud will close again from the same direction.