Surface Laptop Ultra edges an M4 Max Mac Studio on local models, plugged in
Microsoft's $4,300 Surface Laptop Ultra ran local models generally a bit faster than an M4 Max Mac Studio, The New Stack's reviewer found. The edge was measured on the charger with models sized to the machine, so buyers should check their own model mix.
The Engineer · Build desk
Tokens per second on the Surface Laptop Ultra: dense Qwen 3.8 27B at 8-bit, no more than 8 or 9; Qwen 3.8 27B at 4-bit, 25 to 28; 8-bit Gemma 4 26B-A4B, around 40; Nemotron 3.5 Bolt, 56 to 60, sometimes more.
- no more than 9 tokens/sec Qwen 3.8 27B, dense 8-bit, claim 4 reported as 8 or 9 in Unsloth Studio
- at least 25 tokens/sec Qwen 3.8 27B at 4-bit, claim 5 up to 28 tokens per second
- around 40 tokens/sec Gemma 4 26B-A4B at 8-bit, claim 7 generally hit this speed
- at least 56 tokens/sec Nemotron 3.5 Bolt, claim 3 up to 60, sometimes more
| Figure | Value | Note | Claim |
|---|---|---|---|
| Qwen 3.8 27B, dense 8-bit, no more than | 9 tokens/sec | reported as 8 or 9 in Unsloth Studio | 4 |
| Qwen 3.8 27B at 4-bit, at least | 25 tokens/sec | up to 28 tokens per second | 5 |
| Gemma 4 26B-A4B at 8-bit, around | 40 tokens/sec | generally hit this speed | 7 |
| Nemotron 3.5 Bolt, at least | 56 tokens/sec | up to 60, sometimes more | 3 |
What happened
- Nemotron 3.5 Bolt, a heavily quantized model with 9B active parameters, answered at 56 to 60 tokens per second on the laptop.
- The dense Qwen 3.8 27B at 8-bit never passed 8 or 9 tokens per second in Unsloth Studio, and a 4-bit build reached 25 to 28.
- The 8-bit Gemma 4 26B-A4B, a mixture-of-experts model, generally ran at around 40 tokens per second.
- On battery, the 4-bit Qwen fell to around 18 tokens per second and Nemotron to 42 to 45.
Why it matters
- Who has to decide decision Whether this is an option next to a Mac Studio depends on the model mix: 4-bit and mixture-of-experts models run well here, 8-bit dense models do not.
- What it costs cost Moving from the tested 64GB to 128GB adds $1,600, so the memory choice is also a budget choice for a team buying several units.
- What doesn't add up contradiction Microsoft's 99.6% battery-retention claim does not match what the reviewer saw on AI workloads, so unplugged speed should be tested on your own models.
The comparison rests on one reviewer's week and one reference machine. His Mac Studio has an M4 Max and runs most local models he tries, as long as they fit in 48GB [2]. The Surface he tested has 64GB [1]. His summary: "At no point did I feel like I was making a compromise in running those on the laptop." [14]
The review does not include the Mac token rates behind the "bit faster" verdict [15]. Nor does it include a Mac price, only the framing that the Surface undercuts Apple [23].
Tom's Hardware reports that the Ultra generally outperforms the M5 Max in a MacBook Pro on AI work [21]. The reviewer has no M5 machine to confirm it [21].
The Qwen result shows where the machine is weak. The 27B model fits in memory, yet the reviewer says that with dense models a slow result is "virtually always due to memory bandwidth" [6]. Dropping from 8-bit to 4-bit made it roughly three times faster [12]. His conclusion is that mixture-of-experts is "really what you need to run virtually any model locally at reasonable speeds" [8].
Every speed above was taken on the 140W charger [16]. Unplugged, his own ranges work out to a loss of 28% to 36% for the 4-bit Qwen and about 25% for Nemotron [19][20]. He did not run a formal battery-life test. Microsoft quotes 12 hours of active web use. With local AI running constantly, he would not expect more than a few hours [22].
The 64GB configuration sits in an awkward spot. Most smaller models now are 27B to 30B, and few land in the 32GB to 48GB range [9]. The spare memory goes to KV cache and other applications [9]. Fan noise under load is, in his words, "enough to get a bit on your nerves while you are working" [11].
In our view the evidence supports a narrower claim than a general Mac Studio alternative. For 4-bit and mixture-of-experts models on wall power, measured against a 48GB M4 Max, the Surface is a real option.
What to watch
- Mac Studio token rates for the same models and quantizations, including M5 Max results. Those would size the margin the review leaves open.
- Independent battery tests of AI workloads on the Surface Laptop Ultra against Microsoft's 99.6% claim.
- New model releases in the 32GB to 48GB range. Releases there would make the 64GB configuration easier to justify.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence50
- Adoption
- Not scored
- Hype gap+15
- Incentives
- Not scored
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The reviewer used the Surface Laptop Ultra for a week, including on the road, in its high-end $4,300 GPU and CPU configuration with 64GB of RAM and a 1TB SSD.
- [2]
The reviewer typically uses a Mac Studio with the M4 Max chip at home, which runs most local models he throws at it as long as they fit into 48 GB of RAM.
- [3]
Nemotron 3.5 Bolt, a heavily quantized model with 9B active parameters, answered queries at 56 to 60 tokens per second, sometimes more.
- [4]
The dense Qwen 3.8 27B at 8-bit quantization never achieved more than 8 or 9 tokens per second in Unsloth Studio.
- [5]
Bringing Qwen 3.8 27B down to 4-bit quantization raised the speed to 25 to 28 tokens per second.
- [6]
The reviewer said that with dense models, slow speed is "virtually always due to memory bandwidth."
- [7]
The 8-bit Gemma 4 26B-A4B generally hit around 40 tokens per second on the Surface Laptop Ultra.
- [8]
The reviewer wrote that the mixture-of-experts technique is "really what you need to run virtually any model locally at reasonable speeds."
- [9]
With 64GB of RAM, most smaller models are in the 27B to 30B range, there are few models in the 32-48GB range, and the extra RAM gives headroom for the KV cache and other applications.
- [10]
Going for 128GB of RAM brings the price of the Surface Laptop Ultra to $5,900.
- [11]
The reviewer wrote that the noise from the dual fans is not extreme, but "enough to get a bit on your nerves while you are working."
- [12]
Moving Qwen 3.8 27B from 8-bit to 4-bit made it roughly three times faster.
- [13]
The 128GB price of $5,900 is $1,600 above the $4,300 tested configuration.
- [14]
The reviewer wrote: "At no point did I feel like I was making a compromise in running those on the laptop."
- [15]
The reviewer found the Blackwell GPU was generally a bit faster than his Mac Studio.
- [16]
All of the reviewer's speed numbers were taken with the laptop on its 140W charger.
- [17]
Microsoft claims its benchmarks showed the laptop retained 99.6% of its performance on battery, but for the AI workloads the reviewer tested, the drop was far more noticeable.
- [18]
On battery, the 4-bit Qwen 3.8 model dropped from 25-28 to around 18 tokens per second and Nemotron 3.5 Bolt dropped to around 42 to 45 tokens per second; the reviewer expects around a 20% to 25% loss overall on battery.
- [19]
On battery, the 4-bit Qwen 3.8 lost about 28% to 36% of its speed.
- [20]
On battery, Nemotron 3.5 Bolt lost about 25% of its speed.
- [21]
The reviewer has no M5-based MacBook or Mac Studio at hand, but Tom's Hardware reports that the Ultra generally outperforms the M5 Max in a MacBook Pro in AI use cases.
- [22]
The reviewer ran no formal battery life tests; Microsoft says users can expect 12 hours of active web usage, but he would not expect more than a few hours when consistently running local AI workloads.
- [23]
The source describes the Surface Laptop Ultra as having a price that undercuts Apple, without giving a Mac Studio price.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- thenewstack.ioRunning local AI models on Microsoft’s Surface Laptop Ultra: What we found
1 article · October 16, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.