What AI Models Fit on Your Laptop: VRAM, RAM, and Memory Architecture, Measured

What you can run on a laptop is decided not by how big the GPU is, but by the "memory architecture." There are three branches: ==Apple unified memory (RAM = VRAM in one pool), a gaming laptop's dedicated VRAM, and a thin laptop's iGPU/CPU. The measured rule is the same - about 0.5GB per billion parameters at 4-bit plus 20-30% overhead. But on laptops there is a twist: a 32GB MacBook runs a 33B model that a 16GB-VRAM gaming laptop cannot (unified memory merges RAM and VRAM into one pool). The catch is Apple's lower bandwidth (measured M4 Max 546GB/s vs RTX 4090 1008GB/s), so tokens are slower==.

In plain terms: laptop memory is a kitchen layout. A Mac (unified) is "one wide island" that spreads out big dishes but works slowly; a gaming laptop (dedicated VRAM) is "a small but fast dedicated counter" that is quick only on what fits; a thin laptop (CPU) is "a portable burner" good only for small things.

Seeing how the three architectures use memory differently makes the rest click.

What decides what runs on a laptop?#

Whether the whole model fits in fast memory. AI inference needs the entire model resident and is bottlenecked by memory bandwidth, not compute. Rough sizing (Q4): 8GB for 7-8B, 24GB for 32B, 48-64GB for 70B. On top, the KV cache grows with context - going 8K to 32K alone can eat another 40% of available memory. So it is not "it's N billion, should be fine" but a calculation including precision and context.

Unified-memory Mac vs dedicated-VRAM gaming - what differs?#

Capacity vs speed. A Mac shares one memory across CPU and GPU, so it loads big models with no copying - a $1,799 Mac runs a 33B that does not fit any consumer NVIDIA GPU. But it is slower per token from lower bandwidth, and Ollama uses only about 70% of unified memory as the VRAM budget, so a 16GB Mac has 11-12GB usable. Gaming is the opposite: about 3x faster when it fits in VRAM, but on overflow it spills to system RAM over PCIe, so a 70B on a 24GB GPU collapses to 3-8 tok/s instead of 30+. Rule: if it fits, NVIDIA; if it does not, unified memory.

What are the laptop-specific traps?#

Three: thermals, shared RAM, and no upgrades. First, thermal throttling - laptops lose performance on sustained decode, and a fanless MacBook Air more so (so benchmark on AC power and expect ±20% variance). Second, shared RAM - the OS and browser use the same memory, so leave headroom (16GB is the minimum for 7B, 32GB is comfortable for context and multitasking). Third, no upgrades - most laptop memory is soldered and cannot be expanded later, so buy generously up front. Advanced Mac users even tune the unified-memory wired limit, but that is not beginner territory.

What fits on your laptop?#

The key is looking at the architecture first and calculating before you buy.

• Capacity, quiet, portable: a high-memory Mac (48GB+ for 30B class, 128GB up to 70B). Accept the latency.
• Speed (models that fit): an RTX 4070+ gaming laptop is fast on 7-32B. Accept the 16GB ceiling, heat, and weight.
• Thin laptop: start with 16GB RAM and a 3-7B Q4 (CPU 5-10 tok/s). Measure exact speed on your own device on AC power.

Note: GB, tok/s, and bandwidth figures are public 2026 measurements and vary by chip, OS, runtime, context, and thermals by ±20% or more (not permanent numbers; some come from commercial guides, so treat as directional). Laptop memory is often soldered, so decide before buying, and measure exact speed on your own device on AC power (leaderboards are only a start). Hardware and models move fast, so this is reviewed quarterly.

Reference links#

• Ollama (local runtime)
• llama.cpp (GGUF inference)
• MLX (Apple Silicon ML framework)
• mlx-examples (Mac LLM examples)
• Gemma 3 1B (light model, Hugging Face)