Best Local LLM for a Laptop (2026): Picks for 8, 16, 32 and 64 GB RAM, and Which Laptop to Buy

The best local LLM for a laptop depends on system RAM when there is no graphics card. With 8 GB, run Qwen3.5-4B or Gemma 4 E2B. With 16 GB, Qwen3.5-9B. With 32 GB, gpt-oss-20b or Gemma 4 26B A4B. With 64 GB, Qwen3.6-35B-A3B with a long context. Speed is set by memory bandwidth, so a 9B model manages roughly 10 tokens a second on DDR5.

Most laptops have no discrete GPU. The model then lives in ordinary system RAM and the processor, or the integrated graphics, does the arithmetic. That changes which models make sense and it changes what “fast” means. This guide has two halves: what to run on the laptop you already own, by RAM, and which laptop to buy if you are choosing one for local models.

File sizes below are the GGUF downloads on Hugging Face, hardware figures come from Apple, NVIDIA, AMD, Intel, ASUS and HP product pages, and tool behaviour from the Ollama, llama.cpp and Microsoft documentation. We read each of them on 14 and 15 November 2026. If your laptop does have a graphics card, the guide to model picks by graphics memory is the better starting point.

Quick picks by laptop RAM

These picks assume no discrete GPU, a 4-bit GGUF file and one person chatting. The budget column is our planning figure for the model plus its cache: about 60% of RAM on 8 and 16 GB machines and 75% above that, because the operating system, the browser and the integrated graphics all share the same memory.

Why a local LLM without a GPU is bound by memory bandwidth

To produce one token, the processor has to pass over every weight the model uses for that token. For a dense model that is the whole file. A laptop CPU can do the multiplications quickly enough; what it cannot do is pull gigabytes out of RAM any faster than the memory bus allows. More cores do not help once the bus is saturated, which happens with a handful of threads.

Peak bandwidth is simple arithmetic: channels × transfer rate × 8 bytes. Laptops have two memory channels.

Two practical points follow. A laptop with one memory stick in two slots runs single channel, at half the figures above; adding a matching stick is the cheapest speed upgrade there is. And the gap to a graphics card is a factor of six to seventeen, which is why the same model feels so different on a gaming laptop.

Estimated speeds by memory type

The ceilings below divide 60% of peak bandwidth by the bytes read per token, the assumption our tokens-per-second calculator uses. They are estimates for one user, not measurements, and real laptops land at or below them.

Tokens per second. The gpt-oss-20b row comes from the calculator’s mixture-of-experts model with a custom bandwidth entered; the others are the dense arithmetic.

One published measurement shows the estimates are in the right range. Google’s LiteRT-LM card for Gemma 4 E2B, a 2.59 GB file, reports 29.8 tokens a second on an Intel Lunar Lake laptop CPU and 48.4 on the same chip’s integrated GPU.

Best local LLM for 8 GB RAM

An 8 GB laptop has about 4 to 5 GB free once the system and a browser are running. Anything larger than that pushes the model into swap, and a model in swap produces a word every few seconds.

LFM3 is the newest of these: Liquid AI published it in late October 2026 and says a 4-bit copy takes about 1 GB. Its card also says it is not recommended for agentic coding or complex multi-step reasoning, which is a fair description of every model in this tier. Use them for rewriting, summarising a page, extracting fields and simple questions. Our comparison of small models by task goes deeper on which one suits which job.

On DDR4 expect roughly 10 tokens a second from a 4B model. That is slower than reading speed but fine for short answers.

Best local LLM for 16 GB RAM

Sixteen gigabytes is the most common laptop configuration and the first where a no-GPU machine gives answers worth keeping.

Qwen3.5-9B is the balance point. On DDR5 the ceiling is about 10 tokens a second, and on soldered LPDDR5x about 15. Gemma 4 12B is noticeably slower because every token reads 7 GB.

gpt-oss-20b is the model people try next and regret. Its 12.11 GB file loads on a 16 GB laptop, but only by squeezing out everything else, and Ollama’s own tag for it is 14 GB.

Best local LLM for 32 GB RAM

With 32 GB the question changes from “what fits” to “what is still fast”. Dense models of 24B to 27B parameters fit, and they crawl: a 19.4 GB file on DDR5-5600 has a ceiling under 3 tokens a second.

Mixture-of-experts models solve this. They keep a large set of weights in RAM but route each token through a small part of it.

This is the tier behind searches for 32GB RAM LLM performance, and the honest summary is: the sparse models above respond at roughly the pace of a 4B to 9B dense model while answering like something much larger. A dense 27B is for overnight batch jobs, not conversation.

64 GB RAM: what the extra memory buys

On ordinary DDR5, 64 GB does not open a new class of model. gpt-oss-120b is 63.39 GB on its own, and a dense 70B model would produce about one token a second. What the extra memory buys is room:

• Longer context. Qwen3.6-35B-A3B can use its full 256K window with memory to spare.
• Higher precision. Gemma 4 26B A4B at Q8_0 is 26.86 GB and still runs at mixture-of-experts speed.
• Two models loaded. A small fast model for autocomplete and a larger one for questions, without reloading either.

The exception is a laptop whose 64 or 128 GB sits on a wide memory bus, covered in the buying section below. There the larger models are both loadable and usable.

How much context is realistic on a CPU

Memory is rarely what limits context on a laptop. Current small models keep their cache small: a 32,000-token conversation with Qwen3.5-9B costs about 1.1 GB. The limit is how long the processor takes to read the prompt.

Google’s Lunar Lake measurement makes the point. The CPU read the prompt at 435 tokens a second and the integrated GPU at 3,751. An 8,000-token document therefore takes about 18 seconds on the CPU before the first word appears, and about 2 seconds on the integrated graphics, for a 2B-class model. A 9B model is several times slower again.

Three rules keep a CPU-only laptop pleasant:

1. Keep interactive prompts to a few thousand tokens. Paste the relevant section, not the whole report.
2. Set the window deliberately. Ollama’s documentation says it defaults to a 4K context below 24 GiB of graphics memory, which includes every machine with no GPU. Raise it with OLLAMA_CONTEXT_LENGTH; 8K to 16K is a sensible range.
3. Reuse the prompt. Runtimes cache a prefix they have already processed, so a long system prompt only costs time once per session.

Running a local LLM with no GPU: settings that matter

With Ollama, set the context once and check where the model landed:

OLLAMA_CONTEXT_LENGTH=8192 ollama serve
ollama run qwen3.5:9b
ollama ps

The PROCESSOR column of ollama ps shows the CPU and GPU split. With llama.cpp, pick the file, the window and the thread count yourself:

llama-server -hf lmstudio-community/Qwen3.5-9B-GGUF:Q4_K_M -c 8192 -t 8

Settings worth knowing:

• Threads. Set -t to the number of performance cores. Extra threads on efficiency cores often slow generation, because the memory bus is already full.
• Integrated graphics. Ollama’s hardware page says additional GPU support on Windows and Linux comes through Vulkan and is on by default when that backend is installed, and its AMD list includes Ryzen AI processors. An integrated GPU shares the same RAM, so it does not raise the bandwidth ceiling much, but it reads prompts far faster than the CPU.
• Power. Plug in. Laptops cut processor power on battery, and long generations drain one quickly.
• File choice. Ollama’s library tags are often larger than the bare GGUF because they bundle the vision encoder: its qwen3.5:9b tag is 6.6 GB. On a tight machine, download the exact file you sized for. Our quantization guide explains what you give up below 4-bit.

Best laptop for local LLMs: what to buy in 2026

If you are choosing a machine, decide which of four classes you need. The difference between them is where the model’s memory lives.

Apple silicon MacBooks

Every MacBook has unified memory that the GPU can use. Apple’s specification pages list the MacBook Air with M5 at 16, 24 or 32 GB and 153 GB/s, with the 13-inch model from $1,299 on its US store, and the 14-inch MacBook Pro with M6 at up to 170 GB/s, M5 Pro at 307 GB/s with up to 64 GB, and M5 Max at up to 614 GB/s with up to 128 GB. That last configuration is one of very few laptops that holds a 120B-class model. Our MacBook Pro M6 sizing guide works through the 16, 24 and 32 GB versions. Model picks for each configuration are in our guide to models on Apple hardware.

Windows laptops with RTX 50-series GPUs

A laptop GPU is the fastest option per gigabyte, and its memory is fixed at purchase. NVIDIA’s laptop page lists:

Note that the laptop RTX 5090 has 24 GB, not the 32 GB of the desktop card, and that the RTX 5070 is sold in both 8 and 12 GB versions, so check the listing. For which model suits each amount of graphics memory, use the tier-by-tier guide; to compare desktop cards, see the GPU buyer’s guide. A model that overflows the card spills into system RAM and drops to the speeds in the first half of this article.

Strix Halo laptops with unified memory

AMD’s Ryzen AI Max+ 395 puts up to 128 GB of LPDDR5x-8000 on a 256-bit bus, about 256 GB/s, shared between the CPU and a large integrated GPU. Two examples show the range. The ASUS ROG Flow Z13 (2025) is a 1.20 kg tablet-style machine listed with 128 GB and a 200 W adapter. HP’s ZBook Ultra G1a 14 offers the PRO version of the chip with up to 128 GB, of which HP says up to 96 GB can be assigned to the GPU.

By the same bandwidth arithmetic, a 128 GB Strix Halo laptop runs gpt-oss-120b at roughly 30 tokens a second and gpt-oss-20b at about 42. These are the Windows and Linux machines for people who want large models and no CUDA requirement. GPU memory limits and driver choices for the chip are in our Strix Halo guide, and our desk-side hardware comparison covers the same chip in desktop form.

RTX Spark and Intel Core Ultra laptops

NVIDIA’s RTX Spark laptops pair an Arm CPU with an RTX GPU over 24 to 128 GB of unified memory on Windows; our RTX Spark guide covers the memory split and prices. Intel’s current Core Ultra chips are the mainstream choice: Intel’s specification for the Core Ultra X7 358H lists up to 96 GB of LPDDR5x-9600 on two channels and an Arc B390 integrated GPU. That is the same bandwidth class as a base MacBook, enough for the 16 and 32 GB picks above at comfortable speeds.

What actually uses an NPU for LLMs today

Microsoft defines a Copilot+ PC by an NPU that performs more than 40 trillion operations per second. The NPU is a low-power accelerator for integer-quantized models. It is not extra memory and it is not a GPU. Software has to target it explicitly:

• Windows itself. Microsoft says the Copilot+ features that ship in Windows are designed to run on the NPU, and it recommends Windows ML, which loads Qualcomm’s or Intel’s execution provider automatically, for developers.
• Foundry Local. Microsoft’s local runtime detects the hardware and picks the best build of each catalogue model, using GPUs and NPUs when available and falling back to the CPU. It exposes an OpenAI-compatible API on Windows, macOS and Linux.
• Lemonade. This Apache 2.0 server, which AMD engineers optimise for Ryzen AI chips, lists NPU backends for XDNA 2 chips: one for Windows and Linux and one for Windows only, alongside llama.cpp on CPU and Vulkan.
• OpenVINO GenAI. Intel documents running language models on its NPUs through this toolkit.

Ollama’s hardware documentation does not mention NPUs. In practice, a Copilot+ badge tells you the laptop has a recent chip with an NPU; it does not change the model picks in this guide, because the NPU draws on the same system RAM. Our article on CPU and NPU inference for on-premises AI covers where these accelerators fit in an organisation.

A low-end PC checklist

For an older or budget machine, in order:

1. Check RAM and whether both slots are filled. Single-channel memory halves speed.
2. Pick a model from the 8 or 16 GB table, at 4-bit.
3. Set an 8K context and a thread count equal to your performance cores.
4. Keep prompts short and plug in the charger.
5. Check the fit before you download anything larger with the VRAM and KV cache calculator, entering your RAM as the memory pool.

How VDF AI fits

VDF AI does not run on, resell or ship with any of the laptops or runtimes above. It is built for organisations and deploys on their own servers, in a private cloud or on an air-gapped network. The bridge from this guide is the point where one person’s laptop model becomes something a team depends on: someone has to decide who hosts the model, which models are approved, who may use them and what gets recorded.

VDF AI Router registers Ollama and custom on-premises deployments alongside any cloud models your policy allows, returns each routing decision with its reason, and has an air-gap mode that restricts routing to local models. The local LLM overview describes the wider deployment picture, and you can talk to us about a pilot.