Pick by the job. Coding agents and cloud models need 16 to 32 GB of RAM and no GPU. Running models on the laptop needs memory the GPU can reach: 12 to 24 GB of VRAM on an RTX laptop, or 32 GB or more of unified memory on a Mac. Training belongs on rented cloud GPUs.
Our top picks are the MacBook Pro 14 (M5 Pro, 64 GB) at $3,699 for local models, the Razer Blade 16 with a 24 GB RTX 5090 at $4,899.99 if you need CUDA, and the MacBook Air 13 (M5, 32 GB) at $1,699 for everything else (apple.com and razer.com, checked October 21, 2026).
We may earn a commission from links on this page. It doesn’t change what we recommend. How we make money
Our top picks if you are in a hurry
- Core 5 210H · 16 GB
- 512 GB storage
- 16 in, 4.08 lb
- M5 · 32 GB
- 10-core GPU
- 512 GB storage
- 13.6 in, 2.7 lb
- Ryzen AI 7 350 · 32 GB
- RTX 5060 Laptop GPU
- 1 TB storage
- 16 in, 4.19 lb
- Ryzen AI 9 HX 470 · 32 GB
- RTX 5070 Laptop GPU
- 1 TB storage
- 16 in, 3.97 lb
- M5 Pro · 64 GB
- 20-core GPU
- 1 TB storage
- 14.2 in, 3.5 lb
- Core Ultra 9 386H · 32 GB
- RTX 5090 Laptop GPU
- 2 TB storage
- 16 in, 4.6 lb
Three kinds of AI development need three different laptops
Only one of the three needs a big GPU in the laptop.
Job 1 needs less than people think. Claude Code lists “4 GB+ RAM, x64 or ARM64 processor” and an internet connection, with no GPU line (Anthropic). Docker Desktop on Windows asks for 8 GB (Docker), so 16 GB is our floor, 32 GB if Docker runs all day. See our best laptop for Claude Code, Cursor and vibe coding.
Job 3 belongs in the cloud. Hugging Face says “Training a 4B parameter model in mixed precision on a batch size of 16 requires roughly 85GB of GPU memory” (Hugging Face). The biggest laptop GPU has 24 GB.
That leaves job 2.
Memory the GPU can reach decides local AI
A model has to sit in memory the GPU can reach, or it slows down sharply. Laptops offer three kinds.
VRAM on an NVIDIA GPU belongs to the graphics chip alone: 24 GB on the RTX 5090 Laptop GPU, 16 GB on the 5080, 12 GB on the 5070 Ti, 12 GB or 8 GB on the 5070, 8 GB on the 5060 and 5050 (NVIDIA). NVIDIA lists a power range for each (95 to 150 W for the RTX 5090) and makers set their own limits, so two laptops with the same GPU can perform differently.
Unified memory is one pool shared by processor and GPU (Apple silicon, AMD Ryzen AI Max+, NVIDIA RTX Spark N1X). The makers differ on the GPU’s share:
- Apple publishes none. MacBook Air (M5): up to 32 GB, 153 GB/s; MacBook Pro M5 Pro: 64 GB, 307 GB/s; M5 Max (40-core GPU): 128 GB, 614 GB/s (Apple).
- AMD says up to 96 GB of a 128 GB Ryzen AI Max+ 395 can go to graphics, enough for “4-bit LLMs/VLMs up to 128B parameters” (AMD, July 2025).
- NVIDIA lists up to 128 GB on the 6144-core N1X, with no GPU share (NVIDIA). StorageReview’s reviewers set the GPU’s share in advance in the NVIDIA app, because Windows doesn’t yet share the pool on the fly; they gave it 64 GB of 128 GB (StorageReview, October 16, 2026).
Unified memory is soldered: you choose it at checkout and keep it.
System RAM on an RTX laptop catches whatever doesn’t fit in VRAM; the model runs, much slower, and ollama ps shows a split like “48%/52% CPU/GPU” (Ollama FAQ).
How much memory does a model need?
Our arithmetic: weights in GB ≈ parameters in billions × bits per weight ÷ 8. An 8B model at 4 bits is about 4 GB on paper; real files run larger, and you need room on top for the context (the KV cache), the app and the operating system.
| Model size | 4-bit file (Q4_K_M) | 8-bit file (Q8_0) | Smallest laptop memory for the 4-bit file (our view) |
|---|---|---|---|
| 8B (Llama 3.1 8B) | 4.92 GB | 8.54 GB | 8 GB VRAM, or a 16 GB Mac |
| 14B (Qwen3 14B) | 9.00 GB | 15.70 GB | 12 GB VRAM, or a 24 GB Mac |
| 32B (Qwen3 32B) | 19.76 GB | 34.82 GB | 24 GB VRAM (RTX 5090 only), or a 48 GB Mac |
| 70B (Llama 3.3 70B) | 42.52 GB | 74.98 GB | No laptop GPU; 64 GB+ unified |
File sizes from bartowski’s GGUF model cards on Hugging Face, linked under Sources (read October 21, 2026). On a Mac, macOS and your apps share the same pool, hence the larger Mac sizes.
Context costs memory too. Ollama says coding tools and agents “should be set to at least 64000 tokens” (Ollama); by our arithmetic from Qwen3 14B’s configuration, that adds about 10.5 GB to its 9 GB of weights. A local coding agent wants a 24 GB GPU or a 32 GB-plus Mac.
One Ollama user with 48 GB of VRAM on two cards saw gpt-oss:20b run almost entirely on the CPU at about 6 tokens per second; a commenter blamed a 128,000-token context too big for either card (Ollama GitHub). More in how much RAM and VRAM you need for a local LLM.
Will your AI tools run? Windows x64, Windows on Arm and macOS
Windows on Intel or AMD runs everything; Macs run everything except CUDA; Windows on Arm is patchy, going by each maker’s documentation.
| Tool | Windows x64 | Windows on Arm | macOS (Apple silicon) |
|---|---|---|---|
| Ollama | Runs: NVIDIA, AMD (ROCm or Vulkan) | Not in Ollama’s docs; amd64 downloads only | Runs: GPU via Metal, macOS 14+ |
| LM Studio | Runs: needs AVX2 | Runs: Snapdragon X Elite listed | Runs: macOS 14+; no Intel Macs |
| PyTorch 2.14 | Runs: CUDA 12.6, 13.0, 13.2 | Runs with limits: CPU builds; NVIDIA’s own CUDA build for RTX Spark | Runs: MPS, macOS 14+; no CUDA |
| ComfyUI | Runs: NVIDIA, AMD, Intel Arc | Desktop app installs on ARM64; GPU support not documented | Runs: PyTorch MPS |
| Docker Desktop | Runs: WSL 2; NVIDIA GPU passthrough | Early Access; no Windows containers | Runs; GPU access not documented |
| Claude Code | Runs | Runs: ARM64 listed | Runs: macOS 13+ |
Sources: each maker’s requirements page, Microsoft’s PyTorch-on-Arm post and NVIDIA’s RTX Spark forum announcement, linked under Sources; all read October 21, 2026.
CUDA means Windows or Linux with NVIDIA. NVIDIA’s CUDA 13.4 notes list Linux and Windows, not macOS (NVIDIA), and Docker documents GPU passthrough only on Windows with WSL 2 and NVIDIA (Docker). If your stack assumes CUDA, a Mac means rewriting or renting.
Windows on Arm is a per-tool check. Ollama’s docs don’t mention Arm, yet StorageReview ran Ollama tests on N1X laptops and says it used native Arm64 builds of open-source tools. Our stance: until Ollama documents Arm, treat it as unsupported. A maker statement is what we look for, like DaVinci Resolve 21.1 naming “NVIDIA RTX Spark” under Windows for Arm (see our DaVinci Resolve laptop guide).
For image generation, see our best laptop for Stable Diffusion and ComfyUI, and how much VRAM Stable Diffusion needs.
Our picks at a glance
| Pick for | Laptop | Memory the GPU can reach | Price (store, date) |
|---|---|---|---|
| Cheapest that works | Lenovo IdeaPad Slim 5i 16 Gen 10 | Integrated; 16 GB RAM | $929.99, lenovo.com, Oct 22, 2026 (sale) |
| Cloud models and coding agents | MacBook Air 13 (M5, 32 GB) | 32 GB unified | $1,699, apple.com, Oct 21, 2026 |
| CUDA with 32 GB in slots | Gigabyte AERO X16 1VH (RTX 5060) | 8 GB VRAM | $1,249.99, Best Buy, Oct 21, 2026 |
| Cheapest 12 GB VRAM | Lenovo Legion 7 16AGP11 (RTX 5070) | 12 GB VRAM | $2,599.99, Best Buy, Oct 21, 2026 |
| Best for local models | MacBook Pro 14 (M5 Pro, 64 GB) | 64 GB unified | $3,699, apple.com, Oct 21, 2026 |
| Most VRAM, best for CUDA | Razer Blade 16 (2026, RTX 5090) | 24 GB VRAM | $4,899.99, razer.com, Oct 21, 2026 |
| Biggest models | MacBook Pro 16 (M5 Max, 128 GB) | 128 GB unified | $6,999, apple.com, Oct 21, 2026 |
| CUDA on Arm, 128 GB | ASUS ProArt P16 (RTX Spark N1X) | 128 GB unified (64 GB in StorageReview’s setup) | $6,999.99, Best Buy, Oct 21, 2026 (Best Buy showed ‘Coming Soon’ on Oct 22, 2026) |
The best laptop for AI development, by buyer type
Cheapest that works: Lenovo IdeaPad Slim 5i 16 Gen 10
With under $1,000 and an API-and-agents workload, this is what we’d buy: an Intel Core 5 210H, 16 GB of DDR5 (two SO-DIMM slots) and a 16-inch 1920×1200 touchscreen (Lenovo PSREF), sold by Lenovo.
The drawbacks: no GPU for local models, and Lenovo’s sheet lists both “up to 24 GB” and “up to 32 GB”, so plan on 24 GB. Notebookcheck’s roundup cites PCMag finding the display weaker than Dell’s 16 Plus (Notebookcheck).
Best for cloud models and coding agents: MacBook Air 13 (M5, 32 GB)
The laptop we’d hand most developers building on AI. Its 32 GB (Apple) leaves room for Docker, an IDE and a browser, or a 9 GB 4-bit 14B model. This build is $1,699 at apple.com (October 21, 2026).
The drawbacks: no fan, no CUDA, and 32 GB is the ceiling. Notebookcheck found the base M5 Air throttles within a couple of minutes of multi-core load, finishing Cinebench 18% behind a MacBook Pro 14 M5 (Notebookcheck, March 2026).
CUDA with 32 GB in slots: Gigabyte AERO X16 1VH (RTX 5060)
Our pick for CUDA without a 16 GB RAM ceiling: an RTX 5060 Laptop GPU (8 GB), an AMD Ryzen AI 7 350 and 32 GB of DDR5 in two slots that take 64 GB (Gigabyte). Best Buy had it at $1,249.99 on October 21, 2026, up from $1,199.99 three days earlier. Best Buy’s GIGABYTE GAMING A16 pairs an RTX 5060 with 32 GB for slightly less, with a 13th-gen Intel chip and a 1920×1200 screen.
The drawbacks: 8 GB of VRAM holds a 4-bit 8B model but not a 4-bit 14B one. Notebookcheck’s review of the RTX 5070 version found an 85 W GPU limit and loud fans (Notebookcheck, May 2025).
Cheapest 12 GB VRAM: Lenovo Legion 7 16AGP11 (RTX 5070)
12 GB fits 14B models at 4 bits, and this is the cheapest verified laptop we found with it: an RTX 5070 Laptop GPU with 12 GB at 115 W, 32 GB and a 2560×1600 OLED, under 3.97 lb (Lenovo PSREF), $2,599.99 at Best Buy (October 21, 2026).
The drawbacks: the 32 GB is soldered, and Notebookcheck’s review of the RTX 5060 version found the GPU held to 95 W in Performance mode unless set manually (Notebookcheck, April 2026).
Best for local models: MacBook Pro 14 (M5 Pro, 64 GB)
If we were buying one laptop to run local models, this is it. 64 GB holds a 4-bit 70B model (42.5 GB) with room for context, which no laptop GPU can, at 307 GB/s (Apple). Apple offers 64 GB on the M5 Pro only with the 20-core GPU: $3,699 at apple.com (October 21, 2026).
The drawbacks: no CUDA, no documented Docker GPU passthrough, and fixed memory. Notebookcheck put the 20-core M5 Pro between the RTX 5050 and 5060 Laptop GPUs in a 3DMark graphics test (Notebookcheck, March 2026). Our best MacBook for local LLMs compares every configuration.
Most VRAM, best for CUDA: Razer Blade 16 (2026, RTX 5090)
24 GB is as much VRAM as a laptop gets. Razer pairs the RTX 5090 at 165 W with a Core Ultra 9 386H, 32 GB and 2 TB in 4.6 lb (Razer), $4,899.99 on razer.com (October 21, 2026). StorageReview found an RTX 5090 Blade 16 about 2.3 times faster than its RTX Spark review units on models that fit in 24 GB.
The drawbacks: 24 GB stops at 32B models at 4 bits, the 32 GB of system memory is soldered, and Notebookcheck noted a $400 price rise (Notebookcheck, May 2026). See our laptops with the most VRAM.
Biggest models: MacBook Pro 16 (M5 Max, 128 GB)
The most memory the GPU can reach in a laptop without Arm caveats: 128 GB at 614 GB/s, which needs the 40-core M5 Max (Apple). That holds a 4-bit 70B model with plenty of room for context. It’s $6,999 at apple.com (October 21, 2026); the 14-inch with the same chip and memory is $6,699.
The drawbacks: the price, 4.7 lb, no CUDA and fixed memory. Notebookcheck found the M5 Max ahead of RTX 5070 Laptop GPUs in 3DMark and closer in games, but it ran no AI tests (Notebookcheck, March 10, 2026).
CUDA on Arm with 128 GB: ASUS ProArt P16 (RTX Spark N1X)
Our judgment: only for CUDA developers whose tools already run on Windows on Arm. It has the 20-core N1X, 6144-core GPU and 128 GB (ASUS), $6,999.99 at Best Buy (October 21, 2026), though Best Buy’s page said “Coming Soon” on October 22, so check that you can order it. Notebookcheck recommended it for local AI and measured GPU performance on par with the 2025 RTX 5070 ProArt P16 (Notebookcheck, October 19, 2026).
The drawbacks: price, battery life (Notebookcheck’s cons), soldered memory and the Arm gaps below.
For model files: Samsung T7 Shield 1TB
Model files “can be tens to hundreds of GB” (Ollama), and Mac storage is fixed. Samsung rates the T7 Shield at up to 1,050 MB/s over USB 3.2 Gen 2. Samsung’s faster T9 (2,000 MB/s) needs a USB 3.2 Gen 2×2 port none of our picks lists.
NVIDIA RTX Spark (N1X) laptops: buy now or wait?
Wait, unless you need CUDA with 128 GB on Windows and all your tools support Arm. The N1X pairs an Arm CPU with a Blackwell GPU; x64 software runs “through Windows translation”, but plug-ins and drivers may behave differently (Guru3D, October 7, 2026). NVIDIA says CUDA “runs natively on RTX Spark” (NVIDIA). As of October 21, 2026:
- Prices we could find ran from $2,599.99 (Surface Laptop Ultra, microsoft.com) to $6,999.99 (128 GB ProArt P16, Best Buy, which showed it as “Coming Soon” on October 22).
- Small configs are tight. StorageReview’s 32 GB unit, with 12 GB set for the GPU (a split it calls easy to revisit), fell below the 16 GB its smallest Ollama test required, so that unit sat the model tests out; the 128 GB units ran them.
- Python packages lag. An NVIDIA developer-forum announcement installs PyTorch from NVIDIA’s own index and lists packages that lacked Arm64 wheels, such as grpcio and PyArrow (NVIDIA Developer Forums, October 15, 2026).
- Ollama doesn’t name it. Its compute capability 12.1 row lists only NVIDIA’s GB10 (Ollama).
Training: rent the GPU, not the laptop
Write code on the laptop; train on Colab or a rented GPU. Hugging Face’s figures for mixed-precision training sum to 18 bytes per parameter before activations (Hugging Face), so by our arithmetic even a 1B model (18 GB) outgrows a 16 GB laptop GPU.
Colab is free, though “resources are not guaranteed and not unlimited” and free notebooks “can run for at most 12 hours” (Google Colab FAQ). RunPod listed an 80 GB A100 at $1.79 an hour (page updated October 8, 2026; RunPod).
A Mac isn’t a CUDA substitute: one user on the PyTorch Forums found a custom CNN trained considerably slower on a 16 GB M5 Mac than on an RTX 4050 Windows laptop (PyTorch Forums, January 2026).
See our best laptop for machine learning, is a MacBook good for machine learning, how much RAM you need for machine learning and the best laptop for AI and machine learning students. MIT’s 6.S191 labs run “in Google’s Colaboratory” (MIT).
Best laptop for AI development under $1,000
Under $1,000 you’re buying a job-1 laptop, and that’s fine: the IdeaPad Slim 5i 16 above. When we checked the makers’ stores and Best Buy on October 22, 2026, nothing with an RTX 50-series card came in under $1,000 new, and the only new RTX laptop below that price was an HP Victus 15 with a previous-generation RTX 4050 at Best Buy. That’s the price wall our Blender under $1,000 guide hit too. The MacBook Neo ($699 at apple.com, October 18, 2026) runs Claude Code, but its 8 GB is fixed and Docker for Mac alone asks for 4 GB (Docker).
An RTX 4060 or 5060 laptop for machine learning gets its own page.
Laptops to stay away from
- 8 GB laptops for local models. The MacBook Neo’s 8 GB is fixed (Apple), and LM Studio tells 8 GB Mac owners to “stick to smaller models and modest context sizes”.
- Intel Macs. LM Studio doesn’t support them; Ollama runs on them CPU-only.
- Windows on Arm for Docker-heavy or Ollama-first work. Docker on Arm is Early Access; Ollama’s docs don’t mention Arm.
- 24 or 32 GB RTX Spark laptops for local AI, for the reasons above.
- Laptops bought for the “AI PC” sticker. The Copilot+ PC label is about Windows features (see the FAQ), not memory the GPU can reach.
How we chose
We started with each tool’s requirements pages and sorted the work into three jobs. Then we compared memory the GPU can reach from maker spec pages, took US prices on October 18, 21 and 22, 2026, and read StorageReview and Notebookcheck reviews plus owner threads on GitHub and the PyTorch Forums. We haven’t tested these laptops ourselves; how we work explains the method, and our Will it run? checker tests a laptop you already own.
Who we wouldn’t shortlist (yet)
- Alienware 16X Aurora (RTX 5070 Ti). $3,549.99 at dell.com (October 22, 2026) for the Legion 7’s 12 GB of VRAM, at 5.66 lb (Dell).
- Lenovo Legion Pro 7i 16 Gen 10 (RTX 5080, 16 GB). $3,699.99 at lenovo.com (83F500M0US, a sale on October 22, 2026): over $1,000 more than the Legion 7 for 4 GB more VRAM, still short of a 4-bit 32B model (our arithmetic), and 5.67 lb with a 400 W adapter (Lenovo PSREF).
- Lenovo LOQ 15 laptops (RTX 5060 or 5050). Lenovo lists 16 GB as the most memory for the RTX 5060 LOQ Essential 15IRX11 (Lenovo PSREF), and the RTX 5050 LOQ 15IRX10 ships with 16 GB in one stick (Lenovo PSREF); the AERO X16 has twice the RAM for a little more.
- HP ZBook Ultra G1a 14 (128 GB). Up to 96 GB can go to the GPU (HP), and hp.com listed it at about $5,000 on October 22, 2026. We left it out because there’s no CUDA, each tool needs ROCm or Vulkan support, and HotHardware, with 32 GB set aside for graphics, saw under 10 tokens per second from models of about 44 GB. Our local LLM guide covers it.
- RTX 5050 laptops. The same 8 GB as the RTX 5060, and Ollama’s RTX 50 row doesn’t name the 5050, though NVIDIA lists it at compute capability 12.0 and Ollama’s rule is 5.0 or newer (NVIDIA).
Pick your next guide
- Coding with AI tools: best laptop for programming; best laptop for computer science majors.
- Running models: best laptop for running LLMs locally.
- Data: best laptop for data science. Analysts who live in Power BI or Excel need our Power BI guide or heavy Excel guide instead.
FAQ
Do I need an NPU for AI development?
No. Microsoft’s Copilot+ PC label needs an NPU “capable of performing 40+ TOPS”, 16 GB and 256 GB, for Windows features such as Recall and Live Captions (Microsoft). Ollama, LM Studio, PyTorch and ComfyUI document GPU and CPU paths, not laptop NPUs (October 2026). More in do you need an NPU for AI.
Is 16 GB of RAM enough for AI development?
For cloud models and coding agents, yes. For local models, a 16 GB Mac holds 8B models at 4 bits; on an RTX laptop the VRAM matters more, though 32 GB helps when a model spills.
Mac or Windows for AI development?
Mac for the most memory the GPU can reach without Arm caveats; Windows with NVIDIA for CUDA or Docker GPU containers; either for cloud and agent work.
Can I run AI models on a laptop without a GPU?
Yes, slowly. Ollama shows a model loaded “entirely in system memory” as 100% CPU, and ComfyUI’s troubleshooting notes call CPU mode “very slow but works with any hardware”. Aim for small models at 4 bits.
How much VRAM do I need for AI?
8 GB holds 8B models at 4 bits, 12 GB holds 14B with a short context, and 24 GB holds 32B at 4 bits, by our arithmetic from Hugging Face file sizes. 70B models need 64 GB or more of unified memory.
Can a Snapdragon or Copilot+ laptop run Ollama?
Copilot+ laptops with AMD or Intel chips are x64, which Ollama’s Windows downloads cover. On Snapdragon, Ollama’s docs don’t mention Arm; LM Studio does list Snapdragon X Elite. For local models we’d buy x64 or a Mac.
What to do next
- Decide which of the three jobs you’ll do most. If it’s job 1, stop worrying about the GPU.
- Pick the biggest model you want by parameter count, and find memory that holds its 4-bit file plus context.
- Check every tool’s requirements page against your OS and chip, Arm especially.
- Read VRAM from the maker’s spec sheet; on a Mac or RTX Spark laptop, choose memory at checkout.
- Try a free Colab notebook before paying for a bigger GPU.
Sources
Tool requirement pages were read on the date shown; prices were checked on the date in the text.
Tool and platform requirements
- Anthropic: Claude Code advanced setup (system requirements), read October 18 and re-read October 21, 2026 (unchanged)
- Ollama: GPU support, Windows, macOS, context length and FAQ, read October 18 and re-read October 21, 2026 (unchanged)
- LM Studio: system requirements, read October 18 and re-read October 21, 2026 (unchanged)
- PyTorch 2.14 release blog (September 2, 2026), Get Started install page and MPS backend notes, read October 21, 2026
- Microsoft Windows Developer Blog: PyTorch Arm native builds now available for Windows (April 23, 2025), read October 21, 2026
- ComfyUI: system requirements, Desktop for Windows and troubleshooting, read October 21, 2026
- Docker: install Docker Desktop on Windows and on Mac, read October 18 and re-read October 21, 2026 (unchanged); GPU support in Docker Desktop for Windows, read October 21, 2026
- NVIDIA: CUDA Toolkit 13.4 release notes and CUDA GPU compute capability list, read October 21, 2026
- NVIDIA Developer Forums: RTX Spark, getting started with development (an announcement posted by TomNVIDIA, October 15, 2026; the page doesn’t give the poster’s role), read October 21, 2026
- Microsoft: Windows 11 specifications (Copilot+ PC requirements), read October 21, 2026
- Google Colab FAQ, read October 21, 2026
- Hugging Face Transformers: GPU memory usage, read October 21, 2026
- MIT 6.S191 Introduction to Deep Learning labs, read October 21, 2026
- RunPod: GPU pricing (updated October 8, 2026), read October 21, 2026
Memory and model sizes
- NVIDIA: GeForce laptop GPU comparison, read October 18 and 21, 2026
- NVIDIA: RTX Spark specifications and RTX Spark overview, read October 21, 2026
- Apple: MacBook Air tech specs, MacBook Pro tech specs and MacBook Neo tech specs, read October 21, 2026
- AMD: Ryzen AI Max+ 395 and FAQs on Variable Graphics Memory, VRAM and AI model sizes (July 29, 2025), read October 21, 2026
- Hugging Face GGUF model cards by bartowski: Meta-Llama-3.1-8B-Instruct, Qwen3-14B, Qwen3-32B, Llama-3.3-70B-Instruct, read October 21, 2026
Reviews, news and owner reports
- StorageReview: HP OmniBook Ultra 16 review (RTX Spark N1X) (Brian Beeler, October 16, 2026), read October 21, 2026
- Guru3D: RTX Spark brings native CUDA and up to 128GB of unified memory to thin Windows notebooks (October 7, 2026), read October 21, 2026
- Ollama GitHub issue #11731: GPT-OSS:20B running almost entirely on CPU, read October 21, 2026
- PyTorch Forums: Super slow training in M5 Mac compared to 4050 Nvidia GPU (January 8-12, 2026), read October 21, 2026
- Notebookcheck: Lenovo IdeaPad Slim 5 16IRH10 roundup; MacBook Air 13 M5 review (March 7, 2026); Gigabyte Aero X16 review (May 26, 2025); Lenovo Legion 7a 16 G11 review (April 25, 2026); Apple M5 Pro and M5 Max GPU analysis (March 10, 2026; read October 21, 2026); Razer Blade 16 review (May 27, 2026); ASUS ProArt P16 RTX Spark review (October 19, 2026)
Laptop specs and prices
- Lenovo PSREF: IdeaPad Slim 5 16IRH10R and lenovo.com listing (price October 22, 2026)
- Apple: MacBook Air 13-inch M5, 32 GB, MacBook Pro 14-inch M5 Pro, 64 GB, MacBook Pro 14-inch M5 Pro, 48 GB, MacBook Pro 14-inch M5 Max, 128 GB, MacBook Pro 16-inch M5 Max, 128 GB and MacBook Neo (prices October 18-21, 2026)
- Gigabyte: AERO X16 specifications and Best Buy listing (price October 21, 2026)
- Lenovo PSREF: Legion 7 16AGP11 and Best Buy listing (price October 21, 2026)
- Razer: Blade 16 and configurator (price October 21, 2026)
- HP: ZBook Ultra G1a, read October 21, 2026, and hp.com listing, BN2V1UA#ABA (price October 22, 2026); HotHardware: ZBook Ultra G1a 128GB review (August 15, 2025)
- ASUS: ProArt P16 (H7607) tech specs and Best Buy listing (price October 21, 2026; “Coming Soon” on October 22, 2026)
- Microsoft: Surface Laptop Ultra (price October 21, 2026)
- Dell: Alienware 16X Aurora (AC16251), Offer ID ac16251_so_3 (price October 22, 2026)
- Lenovo PSREF: Legion Pro 7 16IAX10H, model 83F500M0US and lenovo.com listing (price October 22, 2026)
- Lenovo LOQ 15: PSREF LOQ Essential 15IRX11 (RTX 5060), read October 21, 2026; PSREF LOQ 15IRX10 (RTX 5050, 83JE0012US) and lenovo.com listing (RTX 5050), read October 22, 2026
- Samsung: Portable SSD T7 Shield 1TB (read October 18, 2026) and Portable SSD T9 2TB (price October 21, 2026)
- Best Buy: HP Victus 15 (15-fb3093dx, RTX 4050), SKU 6623881 (price and seller read October 22, 2026; specs beyond the listing title not confirmed)
- Best Buy: GIGABYTE GAMING A16 (Core i7-13620H, RTX 5060, 32 GB), SKU 6632143 (via our shortlist, price read October 21 and 22, 2026)