Local AI Hardware: RAM, VRAM, and Model Weight Size
A model occupying a few GB may seem to fit if RAM is larger, but operation also needs memory for context, software, and the operating system. Understand the limits before buying hardware.

Start with the task, then choose a model size your machine can handle
No single RAM figure establishes AI suitability. Short classification, long chats, image reading, and concurrent work have different needs. A loadable model may be too slow or unsuitable for Thai. Ask separately: can it load, is it fast enough, and are results usable?
This article focuses on running an existing model to generate answers This covers neither large-model training nor computer-buying recommendations. Drivers, runtime versions, and model formats affect support, so check current documentation before choosing hardware.
RAM, VRAM, and file storage perform different functions.
- Disk space: Stores model files, programs, and multiple copies; it is not all working memory
- RAM: Main memory used by the operating system and programs, including model portions the runtime processes on the CPU
- VRAM: Dedicated graphics card memory for data used by the GPU. VRAM capacity is not the same as the computer's RAM.
- Unified memory: Some architectures let the CPU and GPU share memory. Allow capacity for the whole system rather than adding it up as two separate memory pools.
Having a GPU does not guarantee that the selected program can use the GPU. Ollama hardware support documentation Distinguishes device and backend support. Check actual GPU models, operating systems, and drivers first; search compatibility for the installed model rather than a marketing name.
Estimate raw weight size from parameters and bit precision
For a hypothetical model storing all weights at the same bit depth, start with Weight bytes ≈ parameter count × bits per parameter ÷ 8 These are raw-data counts, excluding quantization scale metadata and working memory. The table calculates a hypothetical 8-billion-parameter model, not actual product minimum specifications.
| Hypothetical weight precision | Calculation | Raw space in GB | About GiB |
|---|---|---|---|
| 16 bits | 8,000,000,000 × 16 ÷ 8 | 16 GB | 14.90 GiB |
| 8 bits | 8,000,000,000 × 8 ÷ 8 | 8 GB | 7.45 GiB |
| 4 bits | 8,000,000,000 × 4 ÷ 8 | 4 GB | 3.73 GiB |
Uses 1 GB = 1,000,000,000 bytes and 1 GiB = 1,073,741,824 bytes, with GiB rounded to two decimals. Actual files may mix precision and metadata; a 4 GB file does not mean 4 GB VRAM suffices for the whole workload.
Hugging Face explains quantization as reducing the precision of numerical representations to save resources, with a trade-off in error. Speed depends on hardware and runtime support. Do not assume fewer bits always mean faster operation or the same quality loss across tasks.
Allow memory for context, KV cache, and other programs
Many language-model architectures retain intermediate values in a KV cache to generate continuous text. Context size, concurrent workloads, and model details therefore affect memory requirements too. Ollama states that increasing the context length increases the memory required. but there is no single formula that applies to every model.
If a model loads but fails with long documents, check configured context and actual memory use before assuming model corruption. Reduce concurrent tasks, close resource-heavy programs, and establish a short-text baseline. Do not maximize context merely because the model supports it.
CPU offloading or disk use instead of RAM can change speed substantially. Inspect runtime and operating-system status rather than diagnosing from RAM size alone.
Test it on the same machine with a small batch before purchasing more.
- Write down CPU, RAM, GPU, VRAM, and free disk space. Includes operating system and runtime version
- Choose a runtime-supported model and check its license and model card for your use
- Begin with one concurrent task and short context. Record time to first response, completion time, and peak observed memory use.
- Use your own Thai tasks, such as public-text summaries, action-item extraction, and insufficient-information questions, with independently checked answers prepared beforehand
- Add documents or context gradually and see whether speed and quality remain acceptable
- Record the model version and weight precision with results. Do not combine measurements from different conditions as one experiment.
If you use Ollama, run this command to check resource allocation and context size.
ollama ps
This check is for your own machine. We did not benchmark performance with this command.
Even with local AI, check where data is sent
Local inference allows data to be entered locally. But the program may have a cloud mode, web search, or additional interfaces. Ollama's FAQ Distinguishes local and cloud use and explains disabling cloud features. Check the model, settings, and called tools rather than inferring local processing from installation alone.
Initial model downloads are separate network activity. Local files and history still need appropriate access controls and backups. For browser-based models, read LiteRT.js and AI in the browser This is another route that has runtime limitations and different devices.
Buy more when you know where the limitations are.
More RAM alone may not fix poor Thai-language quality. If memory is insufficient, identify the bottleneck and whether a smaller model or context is acceptable. If speed is insufficient, check actual accelerator use before buying new hardware.
Before buying hardware, check whether it can complete your actual tasks quickly enough while leaving resources for other applications. Loading a model once or supporting a large parameter count does not establish that the machine will suit everyday use.
Sources
- Hugging Face: Quantization concepts
- Ollama Docs: Context length
- Ollama Docs: Hardware support
- Ollama Docs: FAQ
- huggingface/transformers
Sources reviewed October 14, 2026, Thailand time. Thai examples and prompts are KrobClick proposals, not product tests. Features and settings vary by version/account. Check originals and outputs before real use.