All sources checked 2026-09-28. Prices are US list prices on that date unless stated.
Qwen 3.5 package sizes in Ollama: 9B 6.6GB, 27B 17GB, 35B 24GB, 122B 81GB.
Source: https://ollama.com/library/qwen3.5/tags
The tags table lists qwen3.5:9b 6.6GB, qwen3.5:27b 17GB, qwen3.5:35b 24GB, qwen3.5:122b 81GB, each with a 256K context window and text plus image input.
gpt-oss-20b runs within 16GB; gpt-oss-120b fits on a single 80GB GPU. Source: https://github.com/openai/gpt-oss (also https://huggingface.co/openai/gpt-oss-120b) OpenAI: gpt-oss-120b is built to “fit into a single 80GB GPU (like NVIDIA H100 or AMD MI300X)”; gpt-oss-20b can “run within 16GB of memory”.
gpt-oss-120b: about 117B total parameters, about 5.1B active per token, mixture of experts. Source: https://github.com/openai/gpt-oss OpenAI: “117B parameters with 5.1B active parameters”; the model uses MXFP4 quantization of its MoE (mixture of experts) weights. gpt-oss-20b is 21B total, 3.6B active.
About half a byte per parameter at 4 bits. Source: https://huggingface.co/docs/transformers/main/en/llm_tutorial_optimization Hugging Face gives the rule of thumb of about 4 bytes per parameter in float32 and 2 bytes in bfloat16/float16; 4 bits is a quarter of 16 bits, so about 0.5 bytes per parameter before overhead. The same guide notes quantization “trades improved memory efficiency against accuracy”.
The KV cache grows with context and takes memory.
Sources: https://huggingface.co/docs/transformers/main/en/cache_explanation and https://docs.ollama.com/faq
Hugging Face: with a KV cache, “memory grows linearly” with sequence length. Ollama: required RAM scales with OLLAMA_NUM_PARALLEL × OLLAMA_CONTEXT_LENGTH, and the K/V cache can be quantized to reduce memory use.
The fastest conventional laptop GPU has 24GB: GeForce RTX 5090 Laptop GPU, 24GB GDDR7. Source: https://www.nvidia.com/en-us/geforce/laptops/50-series/ The spec table lists the RTX 5090 Laptop GPU with 24GB GDDR7, 896GB/s, 10,496 CUDA cores.
An RTX 5090 laptop can run its GPU at 175W. Source: https://rog.asus.com/us/laptops/rog-strix/rog-strix-scar-18-2026/spec/ ASUS lists the Strix SCAR 18 (2026) RTX 5090 Laptop GPU at “1647MHz at 175W”, with “150W+25W Dynamic Boost”. NVIDIA’s own page does not give a single power figure because the laptop maker sets it.
RTX PRO 5000 Blackwell Laptop GPU: 24GB GDDR7. Source: https://www.nvidia.com/en-us/products/workstations/professional-laptops/compare/ NVIDIA’s comparison table lists 24GB GDDR7 ECC, 10,496 CUDA cores, 95–175W.
Lenovo ThinkPad P16 Gen 3 offers the RTX PRO 5000, replaceable memory and three drives. Source: https://psref.lenovo.com/syspool/Sys/PDF/ThinkPad/ThinkPad_P16_Gen_3/ThinkPad_P16_Gen_3_Spec.pdf Lenovo PSREF lists “NVIDIA RTX PRO 5000 Blackwell Generation 24GB GDDR7 Laptop GPU”; memory slots “Four DDR5 SODIMM / CSODIMM slots” (up to 192GB non-ECC); storage “Three M.2 slots”, “up to three M.2 2280 Gen 5 Performance SSD; up to 12TB”.
ROG Strix SCAR 18 (2026) with an RTX 5090 starts at $4,999.99, about $5,000. Source: https://rog.asus.com/us/laptops/rog-strix/rog-strix-scar-18-2026/spec/ ASUS Store lists the RTX 5090 (24GB) models at $4,999.99 and $5,299.99. The family’s “starting at $4,299.99” price belongs to the RTX 5080 (16GB) model, G835LWG.
ASUS ROG Flow Z13 (2025): 13.4-inch detachable, Ryzen AI Max+ 395, up to 128GB LPDDR5X. Source: https://rog.asus.com/us/laptops/rog-flow/rog-flow-z13-2025/spec/ ASUS: “13.4-inch 2.5K (2560 x 1600)” display; “AMD Ryzen AI MAX+ 395”; “128GB LPDDR5X 8000 on board”. It is sold as a 2-in-1 gaming tablet with a detachable keyboard. ASUS lists the tablet at 1.20kg (2.65 lb).
Flow Z13 with 128GB: $3,299.99 at the ASUS US store. Source: https://rog.asus.com/us/laptops/rog-flow/rog-flow-z13-2025/spec/ ASUS Store price for GZ302EA-XS99 (Ryzen AI MAX+ 395, 128GB LPDDR5X): $3,299.99. The 32GB and 64GB models list at $2,099.99 and $2,399.99. Press coverage reported a $2,799.99 launch price in 2025.
Flow Z13 weight: 3.51 lb with its keyboard. Source (press, not ASUS): https://fstoppers.com/reviews/review-asus-rog-flow-z13-portable-powerhouse-photographers-creatives-and-coders-701998 ASUS’s own spec page gives the tablet alone at 1.20kg (2.65 lb).
HP ZBook Ultra G1a: Ryzen AI Max PRO, 14-inch mobile workstation, up to 128GB. Source: https://www.hp.com/us-en/workstations/zbook-ultra.html HP: “14” Mobile Workstation PC”, “HP’s thinnest ZBook ever”, “AMD Ryzen AI Max PRO Series”, “Up to 128GB Unified Memory”, and “assign up to 96 GB of memory exclusively to the GPU”.
A 128GB Ryzen AI Max+ system can assign up to 96GB as graphics memory. Source: https://www.amd.com/en/newsroom/press-releases/2025-1-6-amd-announces-expanded-consumer-and-commercial-ai-.html AMD, 6 January 2025: “up to 128GB of unified memory with up to 96GB available for graphics”.
AMD documents models of up to 128 billion parameters running through Vulkan llama.cpp on Windows. Source: https://www.amd.com/en/blogs/2025/amd-ryzen-ai-max-upgraded-run-up-to-128-billion-parameter-llms-lm-studio.html AMD, 29 July 2025: “enable up to 128 billion parameters in Vulkan llama.cpp on Windows … take full use of the 96GB VGM available on a Ryzen AI MAX+ 395 128GB machine”. Its example is Meta’s Llama 4 Scout 109B (17B active) running in LM Studio.
The Ryzen AI Max NPU is rated at “50+” TOPS. Sources: https://www.amd.com/en/blogs/2025/amd-ryzen-ai-max-395-processor-breakthrough-ai-.html and https://www.amd.com/en/products/processors/laptop/ryzen/ai-300-series/amd-ryzen-ai-max-plus-395.html AMD’s March 2025 blog says “50+ peak AI TOPS XDNA 2 NPU”. AMD’s product spec page says “NPU TOPS: Up to 50 TOPS” (overall “Up to 126 TOPS”).
Strix Halo memory bandwidth is 256GB/s. Sources: https://www.amd.com/en/blogs/2025/amd-ryzen-ai-max-395-processor-breakthrough-ai-.html and https://www.amd.com/en/products/processors/laptop/ryzen/ai-300-series/amd-ryzen-ai-max-plus-395.html AMD’s blog: “taking full advantage of the 256 GB/s bandwidth”. The spec page lists “256-bit LPDDR5x” and “LPDDR5x-8000”. At 614GB/s the M5 Max has about 2.4 times that.
M5 Max: up to 128GB unified memory, a 40-core GPU configuration, 614GB/s. Source: https://www.apple.com/macbook-pro/specs/ Apple: M5 Max with 32-core GPU “460GB/s memory bandwidth”; with 40-core GPU “614GB/s memory bandwidth”; memory up to “128GB (M5 Max with 40-core GPU)”.
16-inch M5 Max: up to 22 hours video streaming. Source: https://www.apple.com/macbook-pro/specs/ Apple: 16-inch “Up to 22 hours video streaming” (14-inch M5 Max: up to 20 hours).
16-inch M5 Max, 40-core GPU, 48GB, 2TB: price, and the cost of 128GB. Source: https://www.apple.com/shop/buy-mac/macbook-pro/16-inch-space-black-standard-display-apple-m5-max-chip-18-core-cpu-40-core-gpu-48gb-memory-2tb-storage Apple Store: “Buy for $4,999.00” (standard display, 48GB, 2TB). Memory options: 64GB “+ $400.00”, 128GB “+ $2,000.00”, which makes the 128GB version $6,999.00.
LM Studio supports MLX on Apple silicon, alongside llama.cpp. Source: https://lmstudio.ai/docs/app “LM Studio supports running LLMs on Mac, Windows, and Linux using llama.cpp.” “On Apple Silicon Macs, LM Studio also supports running LLMs using Apple’s MLX.”
Ollama supports macOS. Sources: https://ollama.com/download and https://docs.ollama.com/gpu Download for macOS: “Requires macOS 14 Sonoma or later”. GPU doc: “Ollama supports GPU acceleration on Apple devices via the Metal API.”
Ollama and llama.cpp support AMD through ROCm and Vulkan. Sources: https://docs.ollama.com/gpu and https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md Ollama supports AMD GPUs via ROCm and lists the Ryzen AI Max+ 395 (gfx1151); “Additional GPU support on Windows and Linux is provided via Vulkan”. llama.cpp documents Vulkan and HIP (ROCm, AMD GPU) backends.
RTX Spark: 6,144 CUDA cores (Blackwell), 20-core Grace CPU, up to 128GB unified memory, 1 petaflop FP4. Sources: https://nvidianews.nvidia.com/news/nvidia-microsoft-windows-pcs-agents-rtx-spark and https://www.nvidia.com/en-us/products/rtx-spark/ NVIDIA, 31 May 2026: Blackwell RTX GPU “with 6,144 CUDA cores”, “20-core NVIDIA Grace CPU”, “up to 128GB of unified memory”, “1 petaflop of AI performance” (FP4 on the product page). A smaller variant has 5,120 CUDA cores and an 18-core CPU.
The Grace CPU is Arm based, and the platform runs Windows. Source: https://nvidianews.nvidia.com/news/nvidia-microsoft-windows-pcs-agents-rtx-spark NVIDIA: “MediaTek, a market leader in Arm-based system-on-a-chip designs, collaborated with NVIDIA on the custom CPU design”; the machines are Windows PCs.
Partner laptops arrive in October 2026. Source: https://blogs.nvidia.com/blog/local-ai-ifa-next-gen-agents-nv-pair-rtx-spark/ NVIDIA, 3 September 2026: “NVIDIA RTX Spark Windows PCs Arrive October 2026”, with six OEMs (ASUS, Dell, HP, Lenovo, Microsoft Surface, MSI) shipping in October. NVIDIA had not published pricing when checked.