Local vs hosted AI: where do ChatGPT, Claude, Gemini and Codex fit?
To run an AI model locally, you need downloadable model weights and a compatible runtime. A hosted assistant processes requests on the provider's infrastructure, so you do not need a dedicated AI GPU to use its website or API.
| Product | How it runs | What to calculate |
|---|---|---|
| ChatGPT | Hosted OpenAI assistant. The service runs its models remotely. OpenAI open-weight model guide | No local VRAM estimate for ChatGPT itself. OpenAI gpt-oss is a separate downloadable model family. |
| Claude | Anthropic models accessed through Claude, an API or supported cloud platforms. Anthropic model documentation | The hosted Claude models are not downloadable weights for this calculator. |
| Codex | A coding agent, rather than one fixed model with one memory requirement. OpenAI gpt-oss and Codex local setup | Running a coding tool on your computer does not necessarily run its model locally. With a supported local model setup, calculate memory for that underlying model. |
| Gemini | Google models available through hosted apps and APIs. Google Gemini model documentation | For downloadable Google models, look at Gemma. Gemma is a separate family with its own capabilities. |
Google Gemma: a downloadable local alternative
Our catalog includes Gemma 3 in 1B, 4B, 12B and 27B sizes. The 1B model is text-only. The larger variants support images, but our calculator estimates text inference and does not include vision encoder files or image-processing memory. Gemma is not a local copy of Gemini. Read Google's Gemma 3 model card.
Gemma 3 1B memory · 4B · 12B · 27B
OpenAI gpt-oss: open weights for local reasoning
gpt-oss 20B and 120B use a mixture of experts. Only part of the model activates for each token, but all expert weights still need storage and memory for full GPU loading. Our estimates use rounded MXFP4 download sizes, with runtime and context allowances. These models are separate from ChatGPT. Read OpenAI's open model overview.
gpt-oss 20B memory · gpt-oss 120B memory
What about other models?
The catalog also includes downloadable Llama, Qwen, DeepSeek distill and Yi models. A provider can offer both a hosted service and downloadable models, so check the exact model name and its licence. A similar parameter count does not guarantee similar quality or speed.
Should I choose local or hosted AI?
Local inference gives you control over where model processing happens, but you manage downloads, updates and hardware. Hosted services reduce setup and hardware needs, but depend on provider availability, usage limits and data policies. A local app can still use online tools, so check its settings if keeping data on your computer matters.
Estimate local GPU and RAM needs · Browse all models alphabetically