Science & Tech · AI · 4 hrs ago
What PC hardware you need to run AI models locally
Tools such as Ollama and LM Studio let people run AI chatbots directly on their PCs without an internet connection or subscription, keeping their data on the device.
A typical PC with 16GB of memory can run smaller models with about seven to eight billion parameters, though they are less capable than the leading models behind ChatGPT or Claude.
Running a model closer to those leading systems generally requires at least 32GB of memory and a graphics card with 16GB or more of VRAM, which can be costly.
Compressed versions of models use less space, but the model still has to fit in available memory.
How quickly a model responds depends heavily on memory bandwidth, not just raw processing power, and models that fit in a graphics card’s faster memory usually run faster.
For example, an Nvidia RTX 5090 generated text from OpenAI’s gpt-oss-20b at 205 tokens per second in one test, while a PC’s limits can affect how well it handles other tasks at the same time.
Local AI models can run on ordinary PCs, but the models and their speed depend on available memory and hardware.
A PC with 16GB of RAM can run relatively small models of about seven to eight billion parameters.
Running larger models closer to frontier systems generally requires at least 32GB of memory and a graphics card with 16GB or more of VRAM.
Quantized models use less memory, but models still need to fit in the available memory to run.
Speed depends heavily on memory bandwidth, and a model that fits in a graphics card’s VRAM runs faster than one that spills into system memory.
- Who
- PC users running AI models locally.
- What
- The hardware needed depends on the model’s size and desired speed.
- When
- The article was published on November 9, 2026.
- Where
- On a local PC; the article does not specify a location.
- Why
- Hardware determines which models a PC can run and how quickly they respond.
This story does not have two clearly opposing sides.
This story does not have a timeline yet.
- Small models
- About 7 to 8 billion parameters can run on a PC with 16GB of RAM
- Larger models
- At least 32GB of memory and a graphics card with 16GB or more VRAM
- Llama 3.1 8B
- 4.9GB in Ollama’s library
- Llama 3.3 70B
- 43GB in Ollama’s library
- RTX 5090 speed
- 205 tokens per second on OpenAI’s gpt-oss-20b in LMSYS testing
What people are saying
with just 16 GB of memory
memory-bound