Local AI Report #4 — Best Laptops for Local AI

Issue #4 · August 31, 2026 · Last updated: August 31, 2026 · Jakub Rusinowski

In a laptop, memory bandwidth and memory capacity sit at opposite ends of the price list. The fastest memory you can buy comes in the smallest quantity, the largest comes on the slowest bus, and exactly one part escapes the trade-off.

The number that decides how fast your laptop is

Report #3 established the arithmetic for laptops without a discrete GPU. It does not change when you add one, so this report inherits it:

> tokens/sec = memory bandwidth × efficiency ÷ bytes read per token

Generating a token means reading the model's active weights out of memory, and doing it again for the next token. Bandwidth is the divisor that sets the ceiling. Core count, clock speed and TOPS ratings move the other terms hardly at all.

Two consequences run through everything below.

Bytes read per token is not the model's size. For a dense model they coincide. For a Mixture-of-Experts model they do not: every parameter must be resident, but only the activated experts are read per token. That is why Llama 4 Scout occupies 68.2 GB and still decodes faster than a 32B occupying 22.8 GB. Residency is a capacity question; traffic is a speed question. Our methodology has the workings, and what quantisation costs covers the other half of the trade.

Efficiency is bounded, well below 1.0. No real runtime reads memory at theoretical peak, and anything implying it did is wrong. Later in this report that test throws out most of the published laptop figures we found.

If your laptop has no discrete GPU, this report is not the one you want — read Report #3 instead, which sizes models against system RAM rather than against a graphics card.

The bandwidth ladder

Sorted by the only specification that sets your token rate. "Max memory for models" is what a model may actually occupy, which on unified-memory machines is lower than the memory fitted: macOS caps the GPU's share at roughly 75%, and AMD's Variable Graphics Memory caps the allocatable pool at 96 GB of 128.

Two figures above are derived rather than published, and the arithmetic is here so you can check it:

• RTX 5090 Laptop — 256 bits ÷ 8 = 32 B/clock × 28 Gbps = 896 GB/s
• Dual-channel DDR5-5600 (no discrete GPU) — 2 channels × 8 B × 5600 MT/s = 89.6 GB/s

Three things in that table are worth more than the ordering.

The RTX 5080 Laptop matches the 5090 Laptop's bandwidth exactly. Both run 256-bit GDDR7 at the same speed grade, differing in capacity and shader count rather than memory speed. For decoding tokens they are one part in two sizes, which makes the 5090's premium a *capacity* premium: the 24 GB is what you are buying, not the speed. See what runs on a 5080 and Qwen 3 on it; the 12 GB 5070 Ti drops a tier again, and Gemma 3 on it shows where.

The 12 GB RTX 5070 Laptop buys room, not speed. NVIDIA added it in April 2026 using 3 GB modules on the unchanged 128-bit bus, so both capacities run at 384 GB/s. Useful — 12 GB holds a 14B and 8 GB does not — but nothing gets faster.

The RTX 5060 Laptop is missing, and that is not an oversight. Sources give it both 384 and 448 GB/s on the same bus, and the 448 figure appears to belong to the desktop part, which our own page lists at 448 GB/s. We could not resolve it. A row we would have to guess at is worth less than a gap.

Capacity and bandwidth are on opposite ends of the price list

This is the trade the whole category is built on, and it is the reason a shopping list sorted by price tells you nothing useful.

The fastest memory you can put in a laptop is GDDR7 on a wide bus, and it arrives 24 GB at a time because that is what fits on the package. The largest memory you can put in a laptop is LPDDR5X, and it runs at a quarter of the speed. You do not get to pick both — except once.

Every figure below is estimated, computed by this site's engine at Q4_K_M with an f16 KV cache at 8K of context, and published as a range.

Read the diagonal. The parts that are fast have nothing large to run; the parts that hold something large read it slowly.

• The RTX 5090 Laptop is the fastest thing here and tops out at a 32B. A 70B is not close — it needs 45.7 GB.
• The Ryzen AI Max+ 395 holds a 109B Mixture-of-Experts model comfortably and decodes an ordinary 70B at single digits.
• The M5 Max is the exception: 96 GB for a model *and* 614 GB/s to read it with. Nothing else in a laptop chassis does both.
• The M5 in a MacBook Air will hold a 32B and take most of a minute to write a paragraph with it.

Notice what the MoE rows do to the pattern. Qwen 3 30B-A3B occupies 20.0 GB, nearly as much as Qwen 3 32B at 22.8 GB, and decodes several times faster on the same machine because it reads 3.3B active parameters per token instead of 32.8B. If your machine is bandwidth-poor and capacity-rich — which describes every unified-memory part here — a sparse model is the highest-leverage choice available to you. Build it at Q4_K_M.

The "Needs" column assumes 8K of context. Ask for more and the KV cache grows linearly out of the same budget, which is the mechanism context windows explained covers in full — on a 12 GB card it is the difference between a 14B fitting and not.

For anything not in this table, the VRAM calculator sizes it and the analyzer does it against the machine you own.

What the memory shortage did to this list

You cannot write about laptops in 2026 without writing about DRAM, because the component is the story.

A 64 GB DDR5-5600 kit averaged $191 in August 2025 and $1118 in August 2026 — up 485%. A 128 GB DDR5-6400 kit sells for $3399 against a record low of $329. DDR4, which nothing in this report uses, is up 120–177%. Those are Tom's Hardware's tracked figures, checked 2026-08-30.

The cause is not mysterious. High-bandwidth memory for AI datacentres comes off the same wafers as commodity DRAM, and took 23% of them in 2026 against 8% in 2024. Every HBM ramp subtracts directly from what ends up in laptops.

The effect on finished machines is smaller but real: IDC puts average PC selling prices up 17% across 2026 with shipments down 11.3%, and Dell, Lenovo, HP, Acer and ASUS have all warned of 15–20% list-price rises for the second half of the year.

The Apple window that closed

Apple raised Mac prices twice in 2026, and the difference between the two rises matters more than either one.

On 2026-03-03, alongside the M5 MacBook Air, base prices rose and the memory upgrade tiers did not. For about three months that made Apple's 64 GB and 128 GB upgrades the cheapest high-bandwidth portable memory anyone was selling, because Apple's price list is static while the DRAM market is not. A maxed 16-inch MacBook Pro was $7,349.

On 2026-06-25 that stopped being true. The second rise moved base models 13–15% but upgrades 50–67%, and on the M5 Max specifically the 64 GB and 128 GB upgrades doubled. The same configuration reached $10,149.

We flag it because the March framing is still being repeated and is now three months stale. The arbitrage was real. It is gone. For where this sits in the wider year, see the state of local AI in 2026.

This report publishes no prices, deliberately. Amazon permits a price beside a link into their store only when fetched live from their product API, which this site does not use, and our affiliate policy says so. Separately, no product page was reachable from our build environment, so we read no price off any page. Publishing snippet-scraped figures labelled "as of" would invent precision we do not have.

Every machine here has soldered memory

Short section, because it is a short fact with long consequences.

LPDDR5X is soldered. Apple's unified memory is on the processor package. GDDR7 on a laptop graphics chip is soldered. Not one machine in this report has a memory slot that matters for local AI.

So the usual advice — buy the cheap configuration and add memory later — is not merely suboptimal here, it is impossible. The 16 GB MacBook Air is a 16 GB machine forever.

This inverts how you should spend. On a desktop, memory is the upgrade you defer. On these machines it is the only decision you cannot revisit, and it decides which models you can ever run. Buy the memory first and the storage second: an external SSD is a solved problem and an external memory bus is not.

One caveat worth keeping honest: a laptop with SO-DIMM slots and no discrete GPU *is* upgradable, and adding a second stick doubles its bandwidth. That machine is the subject of Report #3, and if that is what you have, that upgrade is the cheapest performance you will ever buy.

The picks, by what you want to run

Organised by intent rather than by price bracket: the trade-off above means "what is the best laptop" has no answer until you say what you want to run on it. Speed figures are estimates at Q4_K_M and 8K context.

You want the most tokens per second and 24 GB is enough

An RTX 5090 Laptop, 24 GB. The top of the ladder at 896 GB/s, holding a 32B at Q4 with room for context: 65–135 tok/s on an 8B, 21–43 tok/s on a 32B, 110–228 tok/s on Qwen 3 30B-A3B. Nothing larger fits, and that is the deal. Found in the Legion Pro 7i, ROG Strix Scar 18 and Razer Blade 16 — see what runs on it and check a model.

You want to hold a large model and can accept single digits

A Ryzen AI Max+ 395 with 128 GB. 96 GB allocatable to the GPU holds what nothing else in a laptop will: Llama 4 Scout at 68.2 GB runs at 8.7–18 tok/s, because it reads only 17B of active parameters per token. An ordinary 70B also fits, at 2.4–4.9 tok/s — slow enough that you should want it for a reason. See the chip's page, its model list, Llama 4 on it and Qwen 3.5 on it.

The Max+ 392 and Max+ 388 carry the same 40-CU graphics on the same 256-bit bus with fewer CPU cores, so they decode at the same speed. Fewer cores costs prompt processing, not generation.

You want both, and can pay for it

A MacBook Pro 16" with M5 Max. The only part in this report that is not a compromise: 614 GB/s *and* 96 GB for a model. A 70B at Q4 lands at 4.8–9.9 tok/s, a 27B at 8.5–18 tok/s, and an 8B at 32–66 tok/s. It is also the machine whose memory upgrades doubled in June. Details on the M5 Max page, what to run on it, and Llama 4 on it.

You want the cheapest machine that is not a compromise

An M5 Pro with 64 GB, or a mid-configuration Strix Halo. The M5 Pro gives 307 GB/s and 48 GB for a model — 4.9–10 tok/s on a 32B, 2.4–5.0 tok/s on a 70B that only just fits. A Strix Halo holds more and reads it slower. The deciding question is whether you want to run anything above 48 GB at all; if not, the M5 Pro is faster — and Llama 4 on the M5 Pro shows exactly where that ceiling bites. See the M5 Pro page and its model list.

You want a thin laptop and will run small models

A MacBook Air with M5 and 24 or 32 GB. 153 GB/s is modest, but a 14B at Q4 runs at 5.2–11 tok/s and an 8B at 8.9–18 tok/s, which is enough for drafting, summarising and rewriting on a fanless machine. Do not buy the 16 GB one if local models are the point: 16 GB leaves about 12 GB for a model and you will meet the ceiling within a week. What runs on an M5, Gemma 3 on it, and setup on macOS.

Coming from an older machine, the upgrade pages do this comparison directly: MacBook Air M2 8 GB, MacBook Air M3 16 GB, MacBook Pro M4 24 GB and MacBook Pro M4 Pro 48 GB.

TGP is the specification nobody prints, and it matters less than you think

Two laptops can carry the same graphics chip 80 watts apart. A 175 W RTX 5090 in a Legion Pro 7i and a 155–160 W one in a Razer Blade 16 are sold under one name. Almost no listing states the figure, and for gaming it is worth roughly 15% in synthetic tests and 5–10% in real ones.

For generating tokens it is worth close to nothing, and the arithmetic at the top of this report says why. TGP changes how fast the chip computes. It does not change how fast memory streams. Decode is bound by the memory bus, and that bus is the same 256-bit GDDR7 at either wattage: a 95 W and a 175 W RTX 5090 Laptop both read weights at 896 GB/s.

Where TGP does show up:

1. Prompt processing. Reading your input is compute-bound in a way generation is not. A long document, a large codebase, a full RAG context — that is where the extra watts land. 2. Sustained thermals. A thin chassis that boosts to its rated power and then throttles will decode more slowly after ten minutes than in the first thirty seconds. The specification sheet describes the first thirty seconds. 3. Battery. See the limits section.

So find out the TGP before you buy — you are entitled to know which product you are getting — but do not pay for the 175 W version expecting more tokens per second. Buy the wattage for prefill and thermals, and the *bandwidth and capacity* for generation. Reference ranges are in the ladder above; the shipping figure is a per-chassis decision and frequently is not published at all, which is the actual scandal here.

What this will not do for you

The honest limits, in the order you will hit them.

1. No laptop here runs a frontier model. The largest thing that fits in 96 GB at Q4 is a 109B Mixture-of-Experts model. Local inference is a different tool, not a cheaper version of the same one. 2. Thermals throttle sustained generation. Every number here is a short-burst estimate. A thin chassis producing a long answer finishes slower than it started, and no specification sheet says by how much. 3. Battery inference is slow. Laptops drop GPU power substantially on battery. Plan to be plugged in. 4. A used desktop card beats all of this on tokens per dollar. A second-hand RTX 3090 gives 24 GB at roughly 936 GB/s for a fraction of any machine here, and a 4090 more still. Neither fits in a bag. If portability is not genuinely a requirement, you are paying a large premium for it — weigh it against renting, against the running cost, and against whether you need an expensive machine at all. 5. Prompt processing is the slow part and we publish no figures for it. Our time-to-first-token model is wrong by roughly an order of magnitude off NVIDIA hardware. 6. An NPU will not help. The 50-plus TOPS ratings are real and largely irrelevant to generation, which is bandwidth-bound. Current runtimes target the GPU. 7. Memory is soldered, per the section above. This is the limit that lasts longest. 8. We hold no benchmarks of our own for any part here. Every figure is a model output. The next section says how much to trust it.

Should you wait for Gorgon Halo?

The common summary is that AMD Ryzen AI Max+ PRO 495 is a rebadged Strix Halo with a clock bump and no reason to wait. On compute that is fair: 3–4%, a 100 MHz boost to 5.2 GHz, and a Radeon 8065S — 40 CU, RDNA 3.5 that is the existing 8060S at a higher clock. If it were only that, the answer would be no.

It is not only that, and the part everyone skips is the part this site cares about:

Derivation: 256 bits ÷ 8 = 32 B/clock × 8.533 Gbps = 273.1 GB/s.

So it is a rebadge on the axis nobody buying one of these cares about, and a 50% capacity increase on the axis they do. 160 GB allocatable is two thirds more model than the 395 holds, while the sub-7% bandwidth gain means nothing gets meaningfully faster. The ceiling moves; the speed does not.

Wait if you specifically want to hold something the 395 cannot: the capacity jump is the only route to it in a portable machine, and OEM systems were stated for the quarter now ending. Buy now if 96 GB holds what you run — waiting then buys a ceiling you will not use, and a new part arriving into this memory market will not arrive cheap.

Do not wait for the market itself to improve. IDC has the shortage persisting through 2027, Intel points at 2028, and SK Hynix has warned it could run longer still. Buying in 2026 to avoid 2026 memory prices does not work, because the memory is already in the laptop.

Methodology, and how much to trust these numbers

The memory figures are exact arithmetic and we stand behind them. Weights are parameter count times the quantisation's bits per weight. Residency uses total parameters and per-token traffic uses active parameters, which is why the MoE rows behave differently. The tables above are generated by calling the same estimator that produces every figure on the VRAM calculator and every /gpu page, not typed in, so this report cannot contradict them.

One difference is worth naming before you spot it. Every figure here is computed at 8K of context; the /gpu pages quote the same models at the site's 4K default. A longer context means a larger KV cache streamed per token, so the numbers here are slightly lower — an 8B on the Ryzen AI Max+ 395 is 17–35 tok/s at 8K against 18–38 tok/s at 4K on its own page. Same engine, same row, different question. 8K is the honest default for a laptop you actually work on.

The speed figures are estimates and every one is a range. Efficiency is fitted per architecture from our calibration set, and that set contains no measurement of any part in this report — Blackwell laptop silicon, M5-generation Apple chips and Strix Halo are all uncalibrated here. The ranges are wide because our confidence is low, and we would rather show you that than pick a midpoint.

Bandwidth is sourced where it is published and derived where it is not. Two figures in the ladder are derived from bus width × memory clock; both show their arithmetic, and both are labelled in the table itself.

The physics check

Before using any third-party figure we test it against the memory bus:

> implied bandwidth = bytes read per token × tokens per second

Exceed the hardware's peak and the figure describes something the hardware cannot do, whatever was measured. Of 8 published figures: 2 survived, 5 were discarded, 1 had its range top rejected while its floor stood.

Four of the discarded figures are 70B claims about the same chip, published by four different sites. A 70B dense model at Q4 reads 42.26 GB per token. At 614 GB/s the arithmetic ceiling at a realistic efficiency of 0.80 is 11.6 tok/s, and our own estimate for that pairing is 4.8–9.9 tok/s. Every published figure we found for it was between 15 and 32. They cannot all be measurements of the same thing, and none of them can be right.

The fifth failed differently: a claim that the Ryzen AI Max+ 395 runs a 70B at BF16 entirely on its integrated GPU. That is 140 GB of weights on a part holding 128 GB and allocating at most 96 GB to graphics — it fails on capacity before bandwidth is considered.

We are not claiming the survivors are correct, only that they are possible. That is a lower bar than most comparisons clear.

What is deliberately missing

No prices. Not one. Amazon's agreement permits a price beside a link into their store only when fetched live from their product API, which this site cannot use — and independently, no product page was reachable from the environment this was built in, so we read no price off any page. Figures seen in search snippets were deliberately not published. The retailer's page is the only place a current number belongs.

No time-to-first-token, as in Report #3: our prefill constant is wrong by roughly an order of magnitude off NVIDIA hardware.

No RTX 5060 Laptop row, because two credible sources give it two different bandwidths on the same bus.

No GPT-OSS at any size, though it is the obvious model to cite on a 128 GB machine. Our own database marks it dense when it is a Mixture-of-Experts model, which would understate its speed enormously. Better left out than published wrong.

No benchmarks of our own for any part here. Full workings: our methodology, the benchmark set, and the glossary. If you own one of these machines, submitting a measurement is the most useful thing you could do with this page.

Frequently Asked Questions

Related

If you are choosing by memory size: Best LLMs for 12 GB VRAM · Best LLMs for 16 GB VRAM · Best LLMs for 24 GB VRAM · Best LLMs for 48 GB VRAM · Best LLMs for 64 GB VRAM

Set up your platform: Local AI on macOS · Local AI on Windows · Local AI on Linux

Pick by what you need it for: Mac buying guide for local AI · GPU buyer’s guide · VRAM requirements · Understanding quantisation · Mini PCs for local AI

How we got these numbers: Our methodology · Measured benchmarks · Check your own machine · Report #3 — small LLMs on 8 GB and 16 GB laptops

Methodology. Estimates are labeled as estimates; verified figures link to their sources. VRAM is weights + KV cache + framework overhead; speed is a three-term latency model. Both formulas are published at /en/methodology.

← All Reports