Open weights closed most of the gap.
Downloadable systems got near-peer on independent benches. Near-peer is not first.
The state of local AI · 2026
The intelligence moved home.
Open weights. Smaller models that punch up. Quantization that keeps the brain. Your RAM is the gate — not a marketing slide. This page is short on purpose. The calculator is the product.
A 27B from 2026 can beat a 70B from 2024.
Parameter count is a memory bill. It is not a grade.
Memory is the constraint.
Same 128 GB Mac. A much smarter file. Speed is the argument you still have to measure.
Own the weights.
Privacy, a file after the API dies, electricity instead of a token meter. Control is extra work.
A checkpoint is only the start.
Quant + engine decide if it loads. The Hugging Face filename is not a runtime.
No fake demo in this tab.
Nothing here is running a local model. Real inference lives on your machine.
What should I run?
Pick RAM (or a preset) and skill. You get one model, one quant, one engine, and a one-line why. Headline is MODEL — QUANT so 24 GB and 128 GB cannot look like the same file. Every number wears a chip.
FACT vs INTERPRETATION
- FACT
- A requirement or sourced claim. It always includes a URL (Hugging Face file listing, model card, this repo’s JSON, llama.cpp table). No URL → not a FACT.
- FACT SIZE
- Required storage. We always say whether the number is a GGUF file size on disk (FACT from a Hugging Face listing) vs RAM vs VRAM.
- RAM
- System memory (CPU / unified memory on Mac).
- VRAM
- GPU memory. Discrete GPUs load weights into VRAM. Laptops with unified memory share one pool.
- INTERPRETATION my suggestion
- Engine choice, which model is “best”, KV-cache headroom. Not a sourced measurement.
- INTERPRETATION my suggestion
- Bandwidth is how fast that memory can feed the model. LLM decode is often memory-bandwidth bound — a bigger/faster bus (HBM, GDDR, Apple unified) is one of the biggest speed factors, not just whether the file fits. Speed-from-bandwidth stays INTERPRETATION unless we cite a source.
Artificial Analysis
Intelligence Index snapshot — 31 Jul 2026
DeepSeek V4 Flash 0731 · 50
Dated snapshot from the 31 Jul 2026 article (Intelligence Index v4.1, max effort). Not a current official ranking of everything, and not a scrape of the full leaderboard.
31 Jul 2026 article · Live model leaderboard
Accessed 29 Aug 2026. Score reused from this site’s data/laptop-series.json / data/models.json (same URL). No other Artificial Analysis scores are invented here.