There is a specific moment when a technology stops being a spec-sheet flex and becomes something you use: when it fits inside the time you were already wasting.
For local AI on laptops, that moment is this month. Three thin-and-light machines I’ve tested all run a 7–8B parameter model fast enough to summarise a PDF, rewrite a paragraph or search your own notes while the fan stays quiet and the Wi-Fi stays off.
What actually happened
The new generation of mobile chips roughly doubled the sustained neural throughput per watt compared to the 2025 batch. In practice that means a 7B model that used to peg a laptop’s CPU at 100% for twenty seconds now answers in three or four, using a fraction of the battery.
Two more things changed quietly:
- Unified, faster memory. Local inference is memory-bandwidth-bound. LPDDR5X at 8,533 MT/s on a 128-bit bus removes the biggest bottleneck for small models.
- The software stopped fighting you. Runtimes now ship quantised models by default and keep them resident, so the second prompt is dramatically faster than the first.
The numbers that matter
| Workload | 2025 thin-and-light | 2026 thin-and-light |
|---|---|---|
| 7B model, first token | 1.4–2.1 s | 0.4–0.7 s |
| 7B model, tokens/sec | 9–14 | 22–31 |
| Battery used per 10 min of local chat | 6–9% | 2–3% |
| Noise under sustained load | audible | inaudible |
These are my own timings on three retail units, not vendor figures, and they vary by model and by how aggressively the OEM tunes power limits. The pattern is consistent though: local AI went from “a demo you show once” to “a thing you leave running”.
The catch nobody puts in the keynote
Three caveats are worth knowing before you buy anything for this feature:
- RAM is the real spec. 16GB is the floor for a 7–8B model with room left for your actual work. 32GB is the comfortable number and it is the upgrade most worth paying for.
- Quantisation is not free. The 4-bit models that make this possible are noticeably worse at maths and at precise citations. Use them for drafting, not for accounting.
- The “AI PC” badge is meaningless. Benchmarks are not standardised yet, and two laptops with the same badge can differ by 3× in tokens per second.
What to do with this
- If you already own a 2025 machine: nothing. Wait one more generation; the gap will widen.
- If you’re buying now: prioritise 32GB RAM over the faster CPU tier. You will feel the RAM.
- If you handle sensitive documents: this is the first generation where a fully offline workflow is genuinely pleasant rather than a science project.
The interesting question is no longer whether local models are good enough. It is what happens when every laptop quietly has one, and the answer to “can I use AI on this?” becomes “you already are”.



