Personal AI: laptops, workstations and their capabilities (part 2)
Updated: 6 days ago
In the first part of this article, we covered uses for local LLMs and suitable computer architectures, compared their performance, and explained how the processor and memory affect speed.
In this second part, we examine which LLM models fit entirely in a computer's memory and what they can do. We then compare ready-made laptops and workstations on the different architectures manufacturers offer for local LLMs.
Which LLM models suit computers with 32-768GB of memory?
Advertising claims about running models with hundreds of billions of parameters should be read alongside the conditions: AMD advertises 300B on Ryzen AI Halo with 192GB of memory, but at 4-bit quantization. That means roughly 150GB for the weights alone, excluding overhead, while the full BF16/FP16 version needs 4 times as much memory: 600GB! Even more revealing is Framework's example: for a computer with 64GB, the company recommends Qwen3.5-122B-A10B in 3-bit Q3_K_S quantization and lists memory usage of 52.5GB, whereas BF16/FP16 weights alone would require around 244GB. Treat such promises carefully, because aggressive quantization to 2-3 bits can significantly degrade a model's answers and behavior.
This article focuses on computers with at least 128GB of memory for professional work with large models capable of replacing cloud solutions in many work tasks. A computer with less memory can also handle simpler programming tasks. That is why the comparison includes the dense Qwen3.8-27B model.
The main reference for systems with 128GB+ is modern open MoE models Qwen3.8-Flash-Next and GLM-5.3-Flash. Qwen3.8 Flash is a cloud release, while Flash-Next is an open release for local use.
A MoE model uses only some of its parameters for each token, reducing the amount of computation. However, the full set of weights must fit in memory, so a large capacity remains important.
Quantizing weights and the KV cache reduces their size by representing numbers more compactly. The effect on quality depends on the model, task, and degree of compression: Q4 may be suitable for work, while Q8 is often close to the original model's quality. The table lists specific published GGUF versions and the combined size of all weight files for each version.
Models for local use can be found on Hugging Face, which hosts original weights, instructions and ready-made quantized versions. Before downloading, check the author, license, file size and compatibility with your chosen software: llama.cpp usually needs the GGUF format, while MLX on a Mac needs a compatible MLX version. You can also find and download models directly in LM Studio or choose a ready-made package from the Ollama library.
Note that the size of a model's weight files is not its total memory usage: reserve capacity for the OS, working buffers and the KV cache, whose size depends on context length.
If a computer or cluster has enough free memory, it can host several copies of one model or different models for simultaneous use.
On Windows and Linux, an easy way to get started is LM Studio with its graphical interface or Ollama; for more flexible configuration and a local API, use llama.cpp. The same tools are available on Macs with Apple Silicon, along with MLX LM, developed for this platform; GPU acceleration uses Metal. On DGX Spark with DGX OS, an Ubuntu-based Linux distribution, you can run CUDA builds of llama.cpp, or use vLLM following NVIDIA's ready-made instructions to serve multiple requests; the software build or container must support ARM64 and the GB10 GPU.
Memory | Selection conditions | |||
|---|---|---|---|---|
32GB | Q4 + headroom does not fit | Q4 + headroom does not fit | Qwen27B: 32K context fits; 64K is borderline; 128K+ does not fit. | |
64GB | Q4 + headroom does not fit | Q4 + headroom does not fit | Qwen27B: context up to 256K fits. | |
128GB | Q4 + headroom does not fit | Qwen27B: up to 256K. Flash-Next: 32K/64K fits; 128K is borderline. | ||
192GB | Q4 + headroom does not fit | Qwen27B and Flash-Next: up to 256K. | ||
256GB | Context up to 256K fits. | |||
512GB | Context up to 256K. A cluster of 4 nodes needs another 36GB for Qwen27B, 60GB for Flash-Next, and 72GB for GLM Flash: the OS and buffers are needed on every node. | |||
768GB | Context up to 256K. A cluster of 4 nodes needs another 36GB for Qwen27B, 60GB for Flash-Next, and 72GB for GLM Flash: the OS and buffers are needed on every node. |
Comparing models on the same test versions
The figures below are percentages from the developers' publications for the original releases, not results for the selected quantized GGUF versions.
Results also depend on the agent harness and settings: using the same benchmark version does not guarantee identical conditions. For example, Qwen3.8-27B was tested on SWE-bench Pro and DeepSWE 1.1 with the Claude Code agent harness and a 256K context.
Alongside open models, the table provides benchmark values for well-known cloud models as a reference.
Some popular benchmarks:
SWE-bench Pro and DeepSWE show coding performance;
Terminal-Bench 2.1 covers autonomous agent tasks;
GPQA Diamond and HLE with tools cover difficult questions and reasoning, with tools in the latter case.
Model and source | SWE-bench Pro | Terminal-Bench 2.1 | DeepSWE v1.1 | GPQA Diamond | HLE with tools |
|---|---|---|---|---|---|
61.7 | 73.0 | 42.2 | 89.2 | No value | |
62.5 | No value | 58.7 | 91.7 | No value | |
No value | 84.3 | 63.4 | No value | 55.3 | |
No value | 90.6 | 74.2 | 90.9 | 63.9 | |
No value | No value | No value | No value | 64.5 | |
No value | No value | No value | 89.9 | 49.0 | |
No value | 88.8 | 72.7 | No value | 64.5 | |
53.4 | No value | No value | 91.3 | No value | |
No value | 85.0 | 58.0 | No value | 57.9 |
Ready-made computers for local AI models
The tables below list computer families, with a focus on configurations with 128GB of memory or more.
Power requirements are listed separately for each family: adapter wattage and the supported charging protocol.
NVIDIA RTX Spark N1X laptops and workstations
Laptop | Model, chip and memory range | Charging with a USB-C adapter (PD) | Documentation and store |
|---|---|---|---|
❌ | Acer SFF RTX Spark | N1X | up to 128GB | - | |
✅ | ASUS ProArt P16 H7607 | N1X | 32-128GB | PD3.1, 140W | |
✅ | ASUS ProArt P14 H7407 | N1X | 32-128GB | PD3.1, 140W | |
❌ | ASUS ProArt GR1X | N1X | 64/128GB | USB-C EPR PD 3.1(180W); 240W adapter included | |
❌ | Dell Creator Edition Desktop | N1X | up to 128GB | - | |
✅ | Dell XPS 16 Creator Edition | N1X | 128GB | Power Delivery is specified for three USB-C 4.0 ports | |
✅ | HP OmniBook Ultra 16, 16-bk0000 | N1X | 32-128GB | PD3.2 USB-C 140W | |
✅ | HP OmniBook X 14 | N1X | up to 128GB | USB PD 3.2; USB-C GaN adapter 140W | |
✅ | Lenovo Yoga Pro 9n / Pro 9n 15 Gen 11 | N1X | 32-128GB | USB-C PD 3.0; USB-C adapter 140W | |
✅ | Lenovo Yoga 9n 2-in-1 Gen 11 16" | N1X | 32-64GB | USB-C PD 3.0; USB-C adapter 140W | |
✅ | Microsoft Surface Laptop Ultra | N1X | 24-128GB | USB PD 3.1, at least 115W | |
❌ | Microsoft Surface RTX Spark Dev Box | N1X | 128GB | - | |
❌ | MSI EdgeMesa N AI+ | N1X | up to 128GB | USB-C PD IN; adapter 240W | |
✅ | MSI Prestige N16 Flip AI+ | N1X | 32-128GB | PD3.1, 65W(20V3.25A)-140W(28V5A) |
NVIDIA GB10 / DGX Spark workstations
Model, chip and memory range | Charging with a USB-C adapter (PD) | Documentation and store |
|---|---|---|
Acer Veriton GN100 | GB10 | 128GB | Included USB-C 240W power supply | |
ASUS Ascent GX10 | GB10 | 64/128GB | PD3.1 EPR. The datasheet lists an input of up to 180W; the included power supply is 240W | |
CyberGeek DGX Spark Compact AI Workstation | GB10 |128GB | Included 240W USB-C power supply; use the supplied adapter | |
Dell Pro Precision with GB10 (formerly Dell Pro Max with GB10) | 64/128GB | Included USB-C power adapter FCM1253 280W, (output profiles: 5V3A, 9V3A, 15V3A, 20V6.5A, 28V5.89A, 36V5.83A, 48V5.83A); PD revision is not specified. In the 20-48V range, these PD profiles use nonstandard current levels! | |
GIGABYTE AI TOP ATOM | GB10 | 128GB | USB-C PD IN; 240W adapter included; required input power, voltage profile and PD revision are not specified. | |
HP ZGX Nano G1n AI | GB10 | 128GB | Dedicated USB-C input, 240W power supply. | |
Lenovo ThinkStation PGX | GB10 | 128GB | Dedicated USB-C input with USB PD 3.1; supplied 240W power supply | |
MSI EdgeXpert AI | GB10 | 64/128GB | Supplied USB PD 3.1 power supply, 240W, 48V5A | |
NVIDIA DGX Spark | GB10 | 128GB | Supplied USB-C PD3.1 240W power supply | |
PNY NVIDIA DGX Spark | GB10 | 128GB | Power supply included; power and PD mode not specified |
Ryzen AI Max+ 395/495 laptops and workstations
Ryzen AI Max 385/390 and Ryzen AI Max PRO 485/490 models with the Radeon 8050S GPU are not considered because of their "obsolescence".
Laptop | Model, chip and memory range | Charging with a USB-C adapter (PD) | Documentation and store |
|---|---|---|---|
❌ | Acer Veriton RA110 | Max+ 395 | 64/128GB | - | |
❌ | AMD Ryzen AI Halo | Max+ 395 | 128GB | Dedicated USB-C input for charging with the supplied power supply | |
❌ | AMD Ryzen AI Halo | Max+ PRO 495 | 192GB unified memory | USB-C power input | |
❌ | ACEMAGIC F9A | Max+ PRO 495 | 192GB | - | |
✅ | ASUS ProArt PX13 HN7306 | Max+ 395 | 128GB | USB PD 45-100W supported | |
✅ | ASUS ROG Flow Z13 GZ302 | Max+ 395 in the highest version | 32-128GB | USB PD 3.0, 45-100W | |
❌ | Beelink GTR9 Pro | Max+ 395 | 128GB | - | |
❌ | CORSAIR AI Workstation 300 | Max+ 395 | 64/128GB | - | |
❌ | Framework Desktop | Max+ 395: 64/128GB, Max+ PRO 495: 192GB | - | |
❌ | GEEKOM A9 Mega | AMD Ryzen AI Max+ 395 | 128GB shared memory | - | |
❌ | GMKtec EVO-X2 | Max+ 395 | 64-128GB | - | |
❌ | GMKtec EVO-X3 | Max+ 395 | 128GB | - | |
❌ | GMKtec EVO-X5 Pro | Ryzen AI Max+ PRO 495 | 192GB unified memory | - | |
❌ | HP Z2 Mini G1a | Max+ PRO 395 in the highest version | 16-128GB | - | |
✅ | HP ZBook Ultra 14 G1a | Max+ PRO 395 in the highest version | 16-128GB | USB-C 140W | |
✅ | HP ZBook Ultra G3a 16 | Ryzen AI Max+ PRO 495 | up to 192GB unified memory | USB-C input; recommended adapter 180W, alternative 140W | |
❌ | Lenovo ThinkCentre X Ultra | Ryzen AI Max+ PRO 495 | up to 128GB unified memory | - | |
❌ | MINISFORUM MS-S1 MAX | Max+ 395 | 64/128GB | - | |
❌ | MINISFORUM MS-S1 MAX-P495 | Max+ PRO 495 | 192GB unified memory | - | |
❌ | MINISFORUM N5 MAX-P495 | Max+ PRO 495 | 192GB | - | |
❌ | MSI PRO MAX EDGE AI+ 11M | Max+ 395 | 128GB | - |
Apple Silicon laptops and workstations: the two latest generations
For MacBook Pro, this covers M5 Max and M4 Max with a 40-core GPU and 128GB. For Mac Studio, it covers the M5 Max / M5 Ultra generation and the previous M4 Max / M3 Ultra. Apple announced the 512GB M3 Ultra version at launch, but the current specifications page shows a maximum of 256GB.
Laptop | Model, chip and memory range | Charging with a USB-C adapter (PD) | Documentation and store |
|---|---|---|---|
✅ | MacBook Pro 14" (2024) | M4 Max | 36-128GB; 128GB requires a 40-core GPU | PD3.0 (<100W) | |
✅ | MacBook Pro 16" (2024) | M4 Max | 36-128GB; 128GB requires a 40-core GPU | PD3.1 (140W) | |
✅ | MacBook Pro 14" | M5 Max | 36-128GB; 128GB requires a 40-core GPU | PD3.0 (<100W) | |
✅ | MacBook Pro 16" | M5 Max | 36-128GB; 128GB requires a 40-core GPU | PD3.1 (140W) | |
❌ | Mac Studio (2025) | M4 Max | 36-128GB; 128GB requires a 40-core GPU | - | |
❌ | Mac Studio (2025) | M3 Ultra | 96-256GB in the current specification; 512GB at launch | - | |
❌ | Mac Studio | M5 Ultra | 96/256/512GB; 512GB: shipping later in October 2026 | - | |
❌ | Mac Studio | M5 Max | 36-128GB; 128GB requires a 40-core GPU | - | Specifications | Amazon.com(36GB) |
MagSafe 3 on a MacBook is a proprietary connector on the laptop side that receives power from a USB-C charging adapter through the appropriate charging cable.
How to provide reliable power for your computer
Check what is included: some configurations are sold without an adapter, especially in Europe; others include one.
For daily, hours-long use of a local AI model, you need a reliable power supply that can continuously deliver the required wattage without interruptions.
Check voltage, current, protocol, and the permitted cable against the documentation for the specific machine. If the manufacturer requires a proprietary power supply or a non-standard mode, a USB-C connector does not justify replacing it with any high-power charger.
If you plan to connect several devices to a multiport power supply, check the output of a single port and how power is redistributed when other devices are connected.