It's a great idea to deploy large language models locally to save money and protect data privacy!

However, diving into the world of models can be confusing with various parameters and models like 7B, 14B, 32B, 70B... The same model even has so many parameters. Which one should you choose?

And what level is my computer at? Which one can it run?

Don't worry! This article will help you clarify your thoughts and tell you in the simplest way how to choose hardware for deploying large language models locally! I guarantee you won't be confused after reading it!

There is a Hardware Configuration and Model Size Reference Table at the bottom of this article.

Understanding Large Language Model Parameters: What do 7B, 14B, and 32B Represent?

• The meaning of parameters: The numbers 7B, 14B, 32B represent the number of parameters in a large language model (LLM), where "B" stands for Billion. Parameters can be thought of as the "weights" learned by the model during training, which store the model's understanding of language, knowledge, and patterns.
• Number of parameters and model capabilities: Generally speaking, the more parameters a model has, the more complex it is, and theoretically it can learn and store richer information, thereby capturing more complex language patterns and performing more powerfully in understanding and generating text.
• Resource consumption and model size: Models with more parameters also require more computing resources (GPU power), larger memory (VRAM and system RAM), and more data for training and running.
• Small models vs. large models:
  • Large models (such as 32B, 65B or even larger): Capable of handling more complex tasks, generating more coherent and more nuanced text, and may perform better in knowledge Q&A, creative writing, etc. However, they have high hardware requirements and run relatively slowly.
  • Small models (such as 7B, 13B): Consume fewer resources, run faster, and are more suitable for running on resource-constrained devices or in latency-sensitive application scenarios. Small models can also perform well in some simple tasks.
• The choice trade-off: Choosing the size of the model requires a trade-off between the capabilities of the model and hardware resources. More parameters are not necessarily "better"; the most suitable model needs to be selected based on the actual application scenario and hardware conditions.

What Kind of Hardware Do I Need to Run a Local Model?

• Core Requirement: Video RAM (VRAM)

  • Importance of VRAM: When running a large model, the model's parameters and intermediate calculation results need to be loaded into the video memory. Therefore, the size of the video memory is the most critical hardware indicator for running a local large model. Insufficient video memory will cause the model to fail to load, or only allow the use of very small models, and may even seriously reduce the running speed.
  • The Bigger, the Better: Ideally, having a GPU with as much VRAM as possible is best, so you can run larger parameter models and get better performance.

• Second Most Important: System Memory (RAM)

  • The role of RAM: System memory RAM is used to load the operating system, run programs, and supplement video memory. When video memory is insufficient, system RAM can be used as "overflow" space, but it will be much slower (because RAM is much slower than VRAM) and will significantly reduce model operating efficiency.
  • Enough RAM is also important: It is recommended to have at least 16GB or even 32GB or more of system RAM, especially when your GPU video memory is limited. More RAM can help alleviate video memory pressure.

• Processor (CPU)

  • The role of the CPU: The CPU is mainly responsible for data pre-processing, model loading, and some model calculation tasks (especially in the case of CPU offloading). A CPU with good performance can improve the model loading speed and assist the GPU in calculations to a certain extent.
  • NPU (Neural Network Processing Unit): Some laptops are equipped with an NPU (Neural Processing Unit), which is a dedicated hardware for accelerating AI calculations. NPU can accelerate specific types of AI operations, including the reasoning process of some large models, thereby improving efficiency and reducing power consumption. If your laptop has an NPU, that would be a plus, but the GPU is still the core of running a local large model. The support and effect of NPU depend on the specific model and software.

• Storage (Hard Drive/SSD)

  • The role of storage: You need enough hard drive space to store model files. Large model files are usually very large. For example, a quantized 7B model may also require 4-5GB of space, and larger models require tens or even hundreds of GB of space.
  • SSD is better than HDD: Using a solid-state drive (SSD) instead of a mechanical hard drive (HDD) can significantly speed up model loading.

Hardware Priority

1. Video Memory (VRAM) (Most Important)
2. System Memory (RAM) (Important)
3. GPU Performance (Computing Power) (Important)
4. CPU Performance (Auxiliary Role)
5. Storage Speed (SSD is better than HDD)

What if there is no dedicated GPU?

• Integrated Graphics and CPU Operation: If you don't have a dedicated GPU, you can still use integrated graphics (such as Intel Iris Xe) or rely entirely on the CPU to run the model. However, performance will be greatly limited. It is recommended to focus on running 7B or even smaller, highly optimized models and use technologies such as quantization to reduce resource requirements.
• Cloud Services: If you need to run large models but your local hardware is insufficient, you can consider using cloud GPU services such as Google Colab, AWS SageMaker, RunPod, etc.

How to run a local model?

For beginners, it is recommended to use some user-friendly tools that simplify the process of running local models:

• Ollama: Operated through the command line, but installation and use are very simple, focusing on quickly running models.
• LM Studio: The interface is simple and intuitive, supporting model download, model management, and one-click running.

Hardware Configuration and Model Size Reference Table

Slide left and right to see all

X86 Laptops
Integrated Graphics Laptop (e.g. Intel Iris Xe)	Shared System Memory (8GB+ RAM)	8-bit, or even 4-bit quantization	≤ 7B (Extremely Quantized)	* Very basic local running experience, suitable for learning and light experience. * Limited performance, slow reasoning speed. * It is recommended to use 4-bit or lower precision quantized models to minimize video memory usage. * Suitable for running small models such as TinyLlama, etc.
Entry-level Gaming Laptop/Thin & Light Dedicated Graphics Laptop (e.g. RTX 3050/4050)	4-8 GB VRAM + 16GB+ RAM	4-bit - 8-bit quantization	7B - 13B (Quantized)	* Can run 7B models relatively smoothly, and some 13B models can also be run through quantization and optimization. * Suitable for experiencing some mainstream small and medium-sized models. * Note that VRAM is still limited, and running large models will be more difficult.
Mid-to-High-End Gaming Laptop/Mobile Workstation (e.g. RTX 3060/3070/4060)	8-16 GB VRAM + 16GB+ RAM	4-bit - 16-bit (Flexible Choice)	7B - 30B (Quantized)	* Can run 7B and 13B models more comfortably, and has the potential to try models around 30B (requires good quantization and optimization). * You can choose different quantization precisions according to your needs to achieve a balance between performance and model quality. * Suitable for exploring more types of medium and large models.

---

ARM (Apple M Series)
Raspberry Pi 4/5	4-8 GB RAM	4-bit Quantization (or lower)	≤ 7B (Extremely Quantized)	* Limited by memory and computing power, it is mainly used to run extremely small models or as an experimental platform. * Suitable for researching model quantization and optimization technologies.
Apple M1/M2/M3 (Unified Memory)	8GB - 64GB Unified Memory	4-bit - 16-bit (Flexible Choice)	7B - 30B+ (Quantized)	* The unified memory architecture makes memory utilization more efficient. Even M-series Macs with 8GB of memory can run models of a certain size. * Higher memory versions (16GB+) can run larger models and even try models above 30B. * Apple chips have advantages in energy efficiency.

---

Nvidia GPU Computers
Entry-level Dedicated Graphics Card (e.g. RTX 4060/4060Ti)	8-16 GB VRAM	4-bit - 16-bit (Flexible Choice)	7B - 30B (Quantized)	* Similar performance to mid-to-high-end gaming laptops, but desktops have better heat dissipation and can run stably for a long time. * Cost-effective and suitable for entry-level local LLM players.
Mid-range Dedicated Graphics Card (e.g. RTX 4070/4070Ti/4080)	12-16 GB VRAM	4-bit - 16-bit (Flexible Choice)	7B - 30B+ (Quantized)	* Can run medium and large models more smoothly and has the potential to try larger parameter models. * Suitable for users with high requirements for local LLM experience.
High-end Dedicated Graphics Card (e.g. RTX 3090/4090, RTX 6000 Ada)	24-48 GB VRAM	8-bit - 32-bit (or even higher)	7B - 70B+ (Quantized/Native)	* Can run most open-source LLMs, including large models (such as 65B, 70B). * You can try higher bit precision (such as 16-bit, 32-bit) to get the best model quality, or use quantization to run larger models. * Suitable for professional developers, researchers, and heavy LLM users.
Server-level GPU (e.g. A100, H100, A800, H800)	40GB - 80GB+ VRAM	16-bit - 32-bit (Native Precision)	30B - 175B+ (Native/Quantized)	* Designed for AI computing, with ultra-large video memory and extremely strong computing power. * Can run ultra-large models and even perform model training and fine-tuning. * Suitable for enterprise-level applications, large-scale model deployment, and research institutions.

Table Supplement

• Quantization: Refers to reducing the numerical precision of model parameters, such as reducing from 16-bit floating point (float16) to 8-bit integer (int8) or 4-bit integer (int4). Quantization can significantly reduce model size and video memory usage, and accelerate inference speed, but may slightly reduce model accuracy.
• Extreme Quantization: Refers to using very low bit precision quantization, such as 3-bit or 2-bit. It can further reduce resource requirements, but the decline in model quality may be more obvious.
• Native: Refers to running the model at its original precision, such as float16 or bfloat16. You can get the best model quality, but resource requirements are the highest.
• Quantized Parameter Range: "Recommended LLM Parameter Range (after Quantization)" in the table refers to the range of model parameters that the hardware can run smoothly under the premise of reasonable quantization. The actual size and performance of the model that can be run also depend on the specific model architecture, the degree of quantization, software optimization, and other factors. The parameter ranges given here are for reference only.
• Unified Memory: A feature of Apple Silicon chips, where the CPU and GPU share the same physical memory, making data exchange more efficient.