GuideaiDeep read12 min read
The best laptops for running local AI models in 2026
Ahmad JJun 20, 2026Updated Oct 23, 2026
A deep read: the full picture, with the receipts.
The best laptop for running local AI models in 2026, in our ranking, is the 14-inch MacBook Pro with M5 Pro. Apple configures it with up to 64GB of unified memory at 307GB/s, and says every MacBook Pro gives the same performance on battery as plugged in. If your models need more than 64GB, the M5 Max goes to 128GB at 614GB/s; if your tools need CUDA, NVIDIA's new RTX Spark laptops pair an NVIDIA GPU with up to 128GB of unified memory.
Local inference is bound by three things: memory capacity (can the weights fit?), memory bandwidth (how fast can they reach the compute?), and the GPU that does the math. Our guide to choosing a laptop for local AI walks through each. The models people run in 2026 also moved the goalposts: the practical sweet spot is Mixture-of-Experts (MoE) models such as Qwen3-30B-A3B (30.5B parameters in total, 3.3B active per token) and OpenAI's gpt-oss-120b (117B in total, 5.1B active). Speed tracks the active parameters, but memory still has to hold all of them, so this guide sizes every pick by what it can hold.
Shopping for a cheap laptop to learn to code and try small models, rather than one to run 30B and 70B models? That is a different job, covered in the best budget laptops for programming and AI work.
Quick picks:
- Best overall: Apple MacBook Pro 14-inch (M5 Pro), up to 64GB at 307GB/s
- Best starter Mac for local models: Apple MacBook Pro 14-inch (M5, base), up to 32GB at 153GB/s
- Best for CUDA with a big memory pool: an NVIDIA RTX Spark laptop (HP OmniBook Ultra 16, Dell XPS 16 Creator Edition), up to 128GB
- Best for big models on x86 Windows: AMD Strix Halo (Ryzen AI Max+ 395): HP ZBook Ultra G1a or ASUS ROG Flow Z13, up to 128GB
- Best for the largest models on macOS: Apple MacBook Pro (M5 Max, 14- or 16-inch), up to 128GB at 614GB/s
- Best for CUDA on a standard x86 Windows laptop: a laptop with an RTX 5090 (24GB) or RTX 5080 (16GB)
- Best for light local use on Windows on Arm: a Snapdragon X2 Elite laptop
No prices, on purpose. A price copied into an article goes stale while still looking current. This guide quotes a price only where a maker publishes one, with the maker's name on it; the buy link carries the live price. Specs were checked against maker pages on 19 October 2026.
Apple MacBook Pro 14-inch (M5 Pro): Best overall#
The M5 Pro treats system memory as GPU memory. Apple's M5 Pro announcement gives up to 64GB of unified memory at up to 307GB/s, an up-to-18-core CPU and an up-to-20-core GPU with a Neural Accelerator in each core. Apple says the M5 Pro has over 4x the peak GPU compute for AI of the M4 Pro. Peak compute is what prompt processing (prefill) leans on: the wait you feel after pasting a long document into a model. Apple's launch release also says M5 Pro and M5 Max "unleash the full potential of LLMs in LM Studio".
What 64GB holds, using Ollama's own download sizes: gpt-oss:20b (14GB), qwen3:30b (19GB) and Llama 3.3 70B (43GB). Ollama's gpt-oss:120b, at 65GB, does not fit. The 70B fits, but the M5 Pro's 307GB/s is half the M5 Max's 614GB/s, and a dense 70B model reads every weight for every token, so dense 70B is where the Max earns its place.
Two things make it our first pick. Apple states that every MacBook Pro gives the same performance whether it is plugged in or not, and rates the 14-inch M5 Pro at up to 22 hours of video streaming. And on a Mac, Ollama's docs say it uses the GPU through Apple's Metal API and MLX's README lists the GPU among its devices, with no separate GPU driver to install.
Who it's for: Anyone who wants to run Qwen3, Llama, Gemma, DeepSeek, Mistral or gpt-oss locally without fighting CUDA drivers or VRAM limits, and wants the laptop to keep that speed away from a plug.
When to pick something else: If your toolchain needs CUDA, look at an RTX Spark or RTX 5090 laptop. If you need more than 64GB, the M5 Max or Strix Halo.
Apple MacBook Pro 14-inch (M5, base): Best starter Mac for local models#
The base M5 (not Pro) is the entry point to unified memory in a MacBook Pro: 16GB, 24GB or up to 32GB of unified memory at 153GB/s, which Apple puts at nearly 30 percent more than the M4. It has a 10-core CPU, a 10-core GPU with a Neural Accelerator in each core and a 16-core Neural Engine, and Apple rates the 14-inch M5 at up to 24 hours of video streaming. With 32GB it holds gpt-oss:20b (a 14GB download) and qwen3:30b (19GB) with room for context. Llama 3.3 70B (43GB) does not fit. For the rest of Apple's line, see the best Macs for local AI.
Who it's for: Students, developers getting started with local models, or anyone whose daily driver is a small or mid-size model (Gemma, Qwen3-8B, Mistral Small, gpt-oss-20b).
When to pick something else: The moment you regularly run dense 30B-class or 70B models, the base M5's bandwidth and 32GB ceiling will frustrate you. Stretch to the M5 Pro, or to the M5 Max for 70B.
NVIDIA RTX Spark laptop: Best for CUDA with a big memory pool#
NVIDIA said on 17 October 2026 that RTX Spark laptops "are now available", and they change the Windows side of this guide. NVIDIA's RTX Spark chip puts a Blackwell RTX GPU and an Arm-based Grace CPU on one package with shared memory. NVIDIA's spec sheet lists the top laptop version with a 6,144-core GPU, a 20-core CPU, 1 petaflop of AI performance and up to 128GB of LPDDR5X unified memory on a 256-bit bus, and its Windows on Arm porting guide gives the bandwidth as 300GB/s. That is CUDA with more than five times the RTX 5090 Laptop GPU's 24GB. NVIDIA says RTX Spark can run 120-billion-parameter models with up to 1 million tokens of context, and Ollama's gpt-oss:120b is a 65GB download, so a 128GB build holds it.
Read the spec sheet line by line: the 18-core version, with a 5,120-core GPU, tops out at 32GB, which does not hold a 70B model.
The machines: HP's OmniBook Ultra 16 (16-inch 3K OLED, up to 128GB, a 99Wh battery HP rates at up to 17 hours of 1080p video playback) and Dell's XPS 16 Creator Edition (up to 128GB, from 4.45 lb; Dell lists its battery life as pending). NVIDIA also names ASUS, Lenovo, Microsoft and MSI as makers, and its launch post lists RTX Spark laptops starting at $2,399.
The catch is Windows on Arm. NVIDIA says Arm-compatible versions exist for more than 97% of the top 15,000 Python packages, and its developer blog lists the first PyTorch-CUDA build among its Windows on Arm work. Its porting advice is still to ship a native Arm64 build or use Arm64EC, so check that every tool you rely on has an Arm64 version. Independent reviews have only just begun: StorageReview's review of the HP OmniBook Ultra 16 is dated 16 October 2026, and its battery results are still to come. This guide uses the makers' figures, so treat NVIDIA's numbers as NVIDIA's.
Who it's for: Developers who need CUDA and want 70B or 120B-class models on a laptop, and who can live with a brand-new platform.
When to pick something else: If you want the most settled software stack today, the M5 Pro or M5 Max. If you need x86 Windows, Strix Halo or an RTX 5090 laptop.
AMD Strix Halo (Ryzen AI Max+ 395): Best for big models on x86 Windows#
AMD's Ryzen AI Max+ 395 puts up to 128GB of unified memory in an x86 Windows laptop. ASUS's spec sheet for the ASUS ROG Flow Z13 (2025) gives it 16 cores and 32 threads, Radeon 8060S graphics, an NPU rated up to 50 TOPS and 32GB, 64GB or 128GB of LPDDR5X-8000 on board, in a 13.4-inch body that weighs 1.20kg. The HP ZBook Ultra G1a is the 14-inch mobile-workstation version: up to 128GB, and HP says up to 96GB can be assigned to the GPU. HP's own page cites Llama 3.1 70B at Q8 running locally and offline.
That 96GB is the number to size against. Ollama's gpt-oss:120b is a 65GB download, so it fits; Llama 3.3 70B at 43GB fits with room to spare.
The trade-off is bandwidth. AMD lists a 256-bit LPDDR5X-8000 memory interface, which works out to 256GB/s: under half the M5 Max's 614GB/s, and below the RTX Spark's 300GB/s. A dense 70B model reads all of its weights for every token, so that 256GB/s caps its speed. An MoE model reads only its active experts (5.1B of gpt-oss-120b's 117B parameters), which is why MoE is the class this chip suits. On software, AMD's ROCm compatibility matrix names the Ryzen AI Max+ 395 (Radeon 8060S, gfx1151) for Windows and, on Linux, Ubuntu.
Coming next: AMD's Ryzen AI Max+ PRO 495 raises the ceiling to 192GB of LPDDR5X-8533, per AMD's spec page. AMD's 23 July update says "developer platforms featuring Ryzen AI Max PRO 400 will be available in the Ryzen AI Halo and OEM systems later this year"; no laptop with it was on a maker page we read for this update.
Who it's for: People who want to run the biggest open models (gpt-oss-120b, large MoE, or 70B dense) on x86 Windows, and who don't mind some setup.
When to pick something else: If you mostly run models at 30B and below and prefer macOS, the M5 Pro. If you need CUDA, an RTX Spark or RTX 5090 laptop.
Apple MacBook Pro (M5 Max, 14- or 16-inch): Best for the largest models on macOS#
The M5 Max adds the bandwidth the others lack. Apple specifies up to 128GB of unified memory, 460GB/s with the 32-core GPU and 614GB/s with the 40-core GPU, and an 18-core CPU. 614GB/s is twice the M5 Pro's 307GB/s and the highest bandwidth any unified-memory laptop in this guide publishes. 128GB holds Llama 3.3 70B (a 43GB download) and gpt-oss:120b (65GB) with room left for long context, and Apple's release says the M5 Max's bandwidth brings "higher token generation for LLMs".
Configure it carefully. Apple's comparison page lists 128GB only with the 40-core GPU, and the ready-made 32-core models start at 36GB. The linked listing is a 16-inch, 40-core, 128GB build.
Who it's for: Researchers, fine-tuners, and developers who need to run or evaluate full 70B-class models locally, run several large models at once, or work with very long context windows.
When to pick something else: If you don't regularly need 70B-class models, the M5 Pro covers you. If you want 128GB on x86 Windows, Strix Halo; with CUDA, an RTX Spark laptop.
RTX 5090 / RTX 5080 laptop: Best for CUDA on a standard x86 Windows laptop#
If your work is CUDA-native (custom kernels, NVIDIA-specific libraries, or inference that must mirror an NVIDIA cloud deployment) and you want a standard x86 Windows laptop, this is the route. NVIDIA's laptop GPU specs give the RTX 5090 Laptop GPU 24GB of GDDR7 on a 256-bit bus and 1,824 AI TOPS, and the RTX 5080 Laptop GPU 16GB of GDDR7 and 1,334 AI TOPS. NVIDIA lists the 5090 at 95 to 150W and the 5080 at 80 to 150W, set by the laptop maker, so read the GPU wattage on the listing: the same GPU name covers a wide range.
VRAM is the ceiling. Ollama's qwen3:30b (19GB) fits in 24GB but not in 16GB; Llama 3.3 70B (43GB) fits in neither, so part of it has to be offloaded to system RAM and the CPU.
Who it's for: ML engineers tied to NVIDIA's stack (CUDA, TensorRT, DeepSpeed) on x86 Windows, or anyone who needs deployment parity with NVIDIA cloud hardware.
When to pick something else: If you need CUDA and more than 24GB, an RTX Spark laptop goes to 128GB. If running on battery matters, Apple says the MacBook Pro keeps the same performance unplugged.
Snapdragon X2 Elite laptop: Best for light local use on Windows on Arm#
Qualcomm's X2 Elite family comes in 12-core and 18-core versions with an NPU rated at 80 to 85 TOPS, and its product brief gives the memory bandwidth as 152GB/s on the X2 Elite and 228GB/s on the X2 Elite Extreme. ASUS's Zenbook A16 (UX3607) is one example: an X2 Elite Extreme with 48GB of LPDDR5X memory on package, 228GB/s, a 16-inch 3K OLED and a 70Wh battery. That 48GB holds qwen3:30b (19GB) or gpt-oss:20b (14GB). It is Windows on Arm, so the same Arm64 check as RTX Spark applies, and there is no NVIDIA GPU and no CUDA.
Who it's for: Travelers, writers, and developers who mostly use hosted models but want a small or mid-size model on the machine as a fallback.
When to pick something else: For 70B-class models, the Zenbook A16 we read stops at 48GB (Qualcomm's brief lists up to 128GB for the X2 Elite and "128+ GB" for the X2 Elite Extreme, so check the build): look at the M5 Max, Strix Halo or RTX Spark.
How much memory each model size needs#
Capacity decides what fits. The download size is the floor, before the context window, your apps and the operating system take their share. Ollama's own download sizes, against the laptops in this guide:
The MoE models in that table are the reason capacity matters more than it used to. gpt-oss-120b has 117B parameters but only 5.1B are active per token, and Qwen3-30B-A3B has 30.5B with 3.3B active: each token reads only the active parameters, not the whole model, but the whole file still has to fit. Our unified memory calculator and quantization guide go further.
Comparison table#
Why there is no price column. A price copied into a table goes stale while still looking current, and a stale price is worse than none. The buy link shows the live price every time you look.
Can the HP OmniBook Ultra 16 run local Python AI scripts on battery better than a Dell XPS?#
It depends which XPS, and no like-for-like battery test under an AI load was available to read for this update. What the makers' own pages say:
Against the XPS 16 Creator Edition, the HP is the same NVIDIA platform, so a Python script that uses CUDA runs the same way on both; check which RTX Spark version each build has, because the 20-core version goes to 128GB and the 18-core to 32GB. The battery question can't be settled until Dell publishes a figure and someone tests both under load. Against the Intel XPS 16, Dell's 31-hour streaming rating is higher than HP's 17 hours of video playback, but the two are different tests, neither is an AI workload, and CUDA needs an NVIDIA GPU, which Dell's XPS page does not list for that model. On Python itself, NVIDIA says more than 97% of the top 15,000 Python packages have Arm-compatible versions; on the Intel XPS everything runs as native x86.
If AI work on battery is the requirement, Apple states it outright: every MacBook Pro gives the same performance plugged in or not. The HP and Dell pages we read make no statement either way.
Can you run local RAG on a laptop with 2GB of RAM?#
Not usefully, and not on Windows 11 at all: Microsoft's minimum for Windows 11 is 4GB. Local RAG (retrieval-augmented generation: a model answering from your own documents) needs two models, an embedding model to index the documents and a chat model to answer from them. A small pair from Ollama's library is nomic-embed-text (a 274MB download, embeddings only) and qwen3:0.6b (523MB): 797MB of model files before the operating system, the search index and your documents. On a 2GB machine that leaves almost nothing.
For RAG over your own files, our advice is 16GB as the floor, which is also what Microsoft requires of a Copilot+ PC. If budget is the constraint, the best budget laptops for programming and AI work covers 16GB machines for coding and small models; this guide is for running bigger models locally.
How to choose#
1. Fit first, then bandwidth#
Two questions, in order. Can the model fit? (Capacity: total memory has to hold the weights, or you spill to swap and performance collapses.) Then, how fast can it run? (Bandwidth: generating tokens is largely a memory-bandwidth-bound problem.) Unified memory (Apple, RTX Spark, Strix Halo) answers the capacity question; among those laptops, the M5 Max's 614GB/s is the highest bandwidth figure a maker publishes.
2. MoE changed the buying math#
The models that define 2026 local use are Mixture-of-Experts (Qwen3-30B-A3B, gpt-oss-120b, and others). Only a slice of their parameters is active per token (3.3B of 30.5B, and 5.1B of 117B), so speed tracks active parameters while memory must still hold the full model. Practical upshot: buy the capacity to fit these models, and don't assume a "30B" or "120B" label means it runs as slowly as a dense model that size.
3. Unified memory vs discrete VRAM#
For big models, a large unified-memory pool beats a small discrete GPU. A 24GB RTX 5090 Laptop GPU can't hold Llama 3.3 70B's 43GB download; a 64GB or 128GB unified machine can. Until this year that meant choosing between CUDA and capacity. RTX Spark is NVIDIA's answer: CUDA with up to 128GB of unified memory, on Windows on Arm.
4. CUDA matters only if your toolchain requires it#
For inference (running models, not training them), Ollama, LM Studio and MLX each document Apple silicon support: Ollama's macOS docs list Apple M series chips with CPU and GPU support, LM Studio's system requirements name Apple Silicon (M1/M2/M3/M4) and macOS 14.0 or newer, and MLX describes itself as an array framework for machine learning on Apple silicon. AMD's ROCm matrix covers Strix Halo. CUDA is mandatory only if you run custom CUDA kernels, use NVIDIA-specific frameworks, or need parity with NVIDIA cloud hardware. If you do, it is RTX Spark (big memory, Arm) or an RTX 5090/5080 (x86, 24GB or 16GB).
5. Know your model-size target before you buy#
Up to 14B: any pick here works, and the base M5 is plenty. 30B-class MoE: 24GB or more (base M5 with 32GB, M5 Pro, RTX 5090, a 48GB Snapdragon build). 70B dense: 64GB or more, and the M5 Max's 614GB/s for speed. 120B-class MoE: 128GB (M5 Max, a 128GB RTX Spark, Strix Halo). Buy for the models you actually run, not the ones you might run someday.
6. If it doesn't need to be a laptop, look at the desktop options#
Chasing 70B and up, and portability isn't essential? NVIDIA's DGX Spark comes with up to 128GB of unified memory at 273GB/s on a desk, and AMD's Strix Halo chip ships in mini PCs too: see the best mini PCs for local AI inference, the best GPUs for local LLMs, and how to build an AI workstation on a tight budget.
Our picks#
🏆 Top pick: Apple MacBook Pro 14-inch (M5 Pro). Up to 64GB of unified memory at 307GB/s, a Neural Accelerator in every GPU core, and Apple's statement that it keeps the same performance on battery make it this guide's top pick for running models locally.
Frequently asked questions
What is the best laptop for running local LLMs in 2026?
Our pick is the 14-inch MacBook Pro with M5 Pro: up to 64GB of unified memory at 307GB/s, and Apple says it keeps the same performance on battery. If your models need more than 64GB, get the M5 Max with 128GB (614GB/s with the 40-core GPU). If you need CUDA, an NVIDIA RTX Spark laptop goes to 128GB; for CUDA on standard x86 Windows, an RTX 5090 laptop with 24GB.
Which laptops can run AI models locally?
A laptop with 16GB of memory runs small models such as qwen3:8b (a 5.2GB download). For 30B-class MoE models such as qwen3:30b (19GB), you want 24GB or more of VRAM or unified memory. For Llama 3.3 70B (43GB), 64GB of unified memory or more. For gpt-oss:120b (65GB), a 128GB machine: the M5 Max, a 128GB RTX Spark laptop, or Strix Halo with its 96GB GPU share.
Can the HP OmniBook Ultra 16 run local Python AI scripts on battery better than a Dell XPS?
Against the XPS 16 Creator Edition, it is the same NVIDIA RTX Spark platform, and Dell lists that model's battery life as pending, so there is nothing to compare yet. Against the Intel XPS 16, Dell rates its battery higher (31 hours of streaming against HP's 17 hours of video, on different tests), but CUDA needs an NVIDIA GPU, which Dell's XPS page does not list for that model. None of the maker pages we read gives a battery figure under an AI load.
Can I run local RAG on a laptop with 2GB of RAM?
Not usefully. Windows 11 needs 4GB, and even a small RAG pair from Ollama (nomic-embed-text at 274MB plus qwen3:0.6b at 523MB) is 797MB of model files before the operating system and your documents. Plan on 16GB.
Which MacBook Pro is best for prompt processing?
The one with the most GPU cores you can afford. Prompt processing (prefill) is limited by compute rather than bandwidth, and every M5-family GPU core has a Neural Accelerator, so the 40-core M5 Max has the most, then the 32-core M5 Max, the 20-core M5 Pro and the 10-core M5. Apple says the M5 Pro and M5 Max have over 4x the peak GPU compute for AI of the previous generation.
Is an RTX Spark laptop better than a MacBook Pro for local AI?
It does one thing a MacBook Pro can't: CUDA, with up to 128GB of unified memory. On the published figures, its 300GB/s sits just under the M5 Pro's 307GB/s and under half the M5 Max's 614GB/s. It runs Windows on Arm, so check your tools have Arm64 builds, and NVIDIA announced it as available on 17 October 2026; independent reviews have only just begun, so this guide uses the makers' figures.
Can I run local AI models on a Windows laptop without an NVIDIA GPU?
Yes. Ollama's Windows docs say it runs as a native Windows application with NVIDIA and AMD Radeon GPU support, through an AMD ROCm v7 driver stack or a Vulkan-capable Radeon driver, though its Windows GPU table lists only Radeon RX and Radeon PRO cards. AMD's ROCm compatibility matrix lists the Ryzen AI Max+ 395 (Radeon 8060S, gfx1151), with Windows 11 25H2 as the supported Windows version for its APUs. Strix Halo is the notable case: HP says up to 96GB of its 128GB can be assigned to the GPU, four times the RTX 5090 Laptop GPU's 24GB, though its 256GB/s of bandwidth makes MoE models the class it suits.
Does the Neural Engine or the NPU actually run the model?
Not in Ollama or MLX, and it is worth knowing that before a spec sheet sells you one. Ollama's docs say it uses the Mac's GPU through Apple's Metal API, and MLX's README lists the CPU and the GPU as the devices it runs on; neither names the Neural Engine. Reverse engineering of the M1 generation found that the Neural Engine cannot read from the shared memory pool directly: it has to copy data into its own local SRAM first. The same work put the block's roofline ridge point at 162 operations per byte, meaning every byte pulled from memory has to feed at least 162 operations before bandwidth stops being the limiting factor. The engineer who reverse-engineered the driver wrote that Apple originally targeted dense image-processing CNN workloads and that the Neural Engine's dataflow still reflects those assumptions, and noted that macOS itself regularly uses the block only to produce upsampled preview images in Finder. The M5 generation adds a Neural Accelerator inside each GPU core (Apple), so the GPU, where the model already runs, now has that hardware itself. The same question is worth asking of Windows NPUs and the TOPS figures attached to them, which we take apart separately.
Is 16GB of RAM enough for local AI in 2026?
It's the floor. 16GB runs qwen3:8b (a 5.2GB download) with room to spare, and OpenAI says gpt-oss-20b runs within 16GB of memory, which leaves little for anything else. The base M5 MacBook Pro configures up to 32GB, and MoE models need memory for the full model even when only part is active, so if you're buying today, treat 32GB as the sensible starting point.
What is a MoE model, and why does it matter for buying a laptop?
Mixture-of-Experts models activate only a fraction of their parameters per token: Qwen3-30B-A3B uses 3.3B of its 30.5B, and gpt-oss-120b 5.1B of its 117B. Each token reads only the active parameters, which is what makes them quick for their size, but they still need memory to hold every parameter. That's why 2026 buying advice leans toward capacity (fit the whole model) even on machines with modest bandwidth.
Do I need to buy the most expensive config right now?
Not unless your models need it. Buy for the model sizes you run: 32GB covers 30B-class MoE models, 64GB covers a 70B download, and 128GB covers gpt-oss:120b. Apple's unified memory is chosen when you order, and ASUS lists the Flow Z13's memory as on board, so pick the capacity at purchase rather than planning to add it later.
Strix Halo or Apple: which should I get for local AI?
If you want macOS for models up to about 30B, or dense 70B at speed, get Apple (M5 Pro, or the M5 Max for 70B). If you need x86 Windows and want to hold 70B to 120B-class open models (especially MoE), Strix Halo's 128GB, with up to 96GB for the GPU, is the route, with a ROCm software stack and 256GB/s of bandwidth as the trade-offs. --- Our pick for most people running local AI models in 2026 is the 14-inch MacBook Pro with M5 Pro: 64GB at 307GB/s, the same performance on battery by Apple's own statement, and a software stack with no GPU driver to install. Two cases change the answer: if your work needs CUDA and models bigger than 24GB, an NVIDIA RTX Spark laptop is now the place to look, once independent tests back NVIDIA's figures. And if you need dense 70B or 120B-class models on a Mac, get the M5 Max with 128GB.
Sources
- NVIDIA, DGX Sparknvidia.com
- Apple, MacBook Pro technical specifications (M5, M5 Pro, M5 Max)apple.com
- Apple Newsroom, M5 Pro and M5 Max (March 3, 2026)apple.com
- Apple Newsroom, M5 (October 2025)apple.com
- ASUS ROG, Flow Z13 (2025) specificationsrog.asus.com
- HP, ZBook Ultra G1a mobile workstationhp.com
- NVIDIA, GeForce RTX 50 Series laptop GPU comparisonnvidia.com
- Qualcomm, Snapdragon X2 Elite product briefqualcomm.com
- Eileen Yoon, Retrospectively Reverse-Engineering Apple's Neural Engineeiln.github.io
- Apple, MacBook Pro overviewapple.com
- Apple, MacBook Pro model comparisonapple.com
- Apple Store, Buy MacBook Proapple.com
- NVIDIA, RTX Spark specificationsnvidia.com
- NVIDIA, RTX Spark Windows on Arm porting guide: system overviewdocs.nvidia.com
- NVIDIA Newsroom, NVIDIA and Microsoft reinvent Windows PCs with RTX Spark (May 31, 2026)nvidianews.nvidia.com
- NVIDIA Blog, RTX Spark laptops now available (October 17, 2026)blogs.nvidia.com
- NVIDIA Developer Blog, Code, test, and deploy on one machine with RTX Spark (October 16, 2026)developer.nvidia.com
- HP Newsroom, new OmniBook PCs (September 4, 2026)hp.com
- Dell, XPS laptops comparisondell.com
- ASUS, Zenbook A16 (UX3607) tech specsasus.com
- AMD, Ryzen AI Max+ 395 specificationsamd.com
- AMD, Ryzen AI Max+ PRO 495 specificationsamd.com
- AMD Newsroom, AAI 2026 client computing update (July 23, 2026)newsroom.amd.com
- AMD, ROCm compatibility matrixrocm.docs.amd.com
- Qwen, Qwen3-30B-A3B model cardhuggingface.co
- OpenAI, gpt-oss-120b model cardhuggingface.co
- Ollama, qwen3 model libraryollama.com
- Ollama, gpt-oss model libraryollama.com
- Ollama, llama3.3 model libraryollama.com
- Ollama, nomic-embed-text model libraryollama.com
- Microsoft, Windows 11 specificationsmicrosoft.com
- Ollama, GPU support docsdocs.ollama.com
- Apple MLX, GitHub READMEgithub.com
- Ollama, Windows docsdocs.ollama.com
- Ollama, macOS docs (system requirements)docs.ollama.com
- LM Studio, system requirementslmstudio.ai
- StorageReview, HP OmniBook Ultra 16 review (NVIDIA RTX Spark N1X), October 16, 2026storagereview.com
Discussion