Skip to content

How-ToCPU

How to Run a Local LLM Without a Graphics Card (CPU-Only Ollama and LM Studio)

Ollama and LM Studio can run local LLMs entirely on the CPU and system RAM. LM Studio needs an AVX2 CPU on x64 and recommends 16GB of RAM; Ollama needs 4GB for the app plus space for models. Pick a small quantised model and check `ollama ps` shows 100% CPU.

By , Editorial team

3 min read

How to Run a Local LLM Without a Graphics Card (CPU-Only Ollama and LM Studio)

Not every PC has a graphics card with lots of VRAM, and plenty of office PCs and mini PCs have none at all. You can still run a local large language model (LLM): both Ollama and LM Studio can load models into system memory and run them on the CPU. This guide is based on Ollama's and LM Studio's official documentation; Forge PC hasn't benchmarked CPU-only speeds, so we don't quote any.

What you need

LM Studio lists these Windows requirements:

  • CPU: AVX2 instruction support is required on x64 processors (Snapdragon X Elite ARM PCs are also supported)

  • RAM: at least 16GB recommended, because LLMs can consume a lot of RAM

  • GPU: at least 4GB of dedicated VRAM is recommended, not required

Ollama runs as a native Windows application. Its Windows documentation says the install needs at least 4GB of space for the program, plus extra space for models, which can be tens to hundreds of GB. If your C: drive is tight, our guide to moving Ollama models to another drive explains how to change where models are stored.

Step 1: install Ollama or LM Studio

Install either app as normal. Our guide to running AI models locally with Ollama and LM Studio walks through the first run. Neither needs a GPU to start.

Step 2: choose a small model

Without a GPU, the whole model sits in system RAM, alongside Windows and your other apps. Start with a small model and a quantised version (for example a 4-bit GGUF in LM Studio) rather than the biggest one you can find. Our GGUF quantisation explainer covers what the labels mean, and how much RAM you need for local AI covers sizing.

Step 3: check where the model is running

Ollama's FAQ says the ollama ps command shows loaded models and a Processor column:

  • 100% CPU means the model was loaded entirely into system memory

  • 100% GPU means it was loaded entirely onto the graphics card

  • a split such as 48%/52% CPU/GPU means it's partly in each

On a PC without a supported GPU you should see 100% CPU. If you do have a GPU but it isn't being used, check Ollama's hardware support page: it lists NVIDIA cards with compute capability 5.0 and newer (driver 550 or newer, or 570+ for compute capability 5.0-6.2), and its Windows page says AMD Radeon cards can use ROCm or Vulkan acceleration.

Step 4: keep expectations realistic

Running on the CPU works, but LM Studio still recommends at least 4GB of dedicated VRAM, so don't expect GPU-like responsiveness (we haven't measured CPU-only speeds). A longer context also uses more memory: see context length and memory explained. Close heavy apps, keep the context length modest (see how to set Ollama's context length), and use a smaller model if replies are too slow.

When to add a GPU

If you find local AI useful, a graphics card with more VRAM is the biggest upgrade. Our guides to running a local LLM on 8GB of VRAM and NPU vs GPU for local AI explain the options.

Our verdict

You can run a local LLM with no graphics card: install Ollama or LM Studio, pick a small quantised model, and confirm with ollama ps that it shows 100% CPU. LM Studio recommends 16GB of RAM and needs an AVX2 CPU on x64; leave plenty of disk space for models.

The bottom line

You can run a local LLM with no graphics card: install Ollama or LM Studio, pick a small quantised model, and confirm with `ollama ps` that it shows 100% CPU. LM Studio recommends 16GB of RAM and needs an AVX2 CPU on x64; leave plenty of disk space for models.

Sources

  1. 01LM Studio, System Requirements
  2. 02Ollama, Windows documentation
  3. 03Ollama, FAQ
  4. 04Ollama, Hardware support

About the author

Forge PC Staff

Editorial team

The shared byline for articles researched, written and edited by the Forge PC editorial team. How we source, check and correct our work is set out in our editorial policy.

More from Forge PC Staff

Read next

How to Install the ARCTIC Freezer 36 CPU Cooler (AM5, AM4, LGA1851 and LGA1700)

How-To

How to Install the ARCTIC Freezer 36 CPU Cooler (AM5, AM4, LGA1851 and LGA1700)

ARCTIC's manual for the Freezer 36 covers AM5, AM4, LGA1851 and LGA1700. On Intel, its contact frame replaces the ILM: support the backplate, keep the ILM parts, align the CPU to the top of the cutout and tighten four screws in a cross pattern. Then choose airflow, connect CPU_FAN and set a fan curve.

Forge PC Staff3 min

How to Check Your Wi-Fi Signal Strength in Windows 11

How-To

How to Check Your Wi-Fi Signal Strength in Windows 11

Check Wi-Fi signal strength in Windows 11 from the taskbar bars, Settings > Network & internet, or exactly with netsh wlan show interfaces (the Signal percentage). Windows Central treats 80-100% as reliable and below 60% as weak. netsh wlan show wlanreport summarises recent wireless sessions.

Forge PC Staff2 min

How to Use Embedding Models in Ollama for Local Semantic Search and RAG

How-To

How to Use Embedding Models in Ollama for Local Semantic Search and RAG

Embedding models turn text into vectors for semantic search and RAG. Pull one such as embeddinggemma in Ollama, generate vectors with ollama run or the /api/embed endpoint, store them, then embed questions with the same model and compare using cosine similarity, all locally.

Forge PC Staff3 min

G.SKILL Trident Z5 Neo RGB vs Trident Z5 RGB: Which DDR5 Kit Should You Buy?

Versus

G.SKILL Trident Z5 Neo RGB vs Trident Z5 RGB: Which DDR5 Kit Should You Buy?

G.SKILL's Trident Z5 Neo RGB is its AMD EXPO kit and the Trident Z5 RGB its Intel XMP kit. The white 32GB Neo is rated DDR5-6000 CL30 at 1.35V with EXPO and XMP; the white Trident Z5 RGB is rated DDR5-6400 CL32 at 1.40V with XMP only. Both work out at about 10ns first-word latency, so choose by platform.

Forge PC Staff3 min