Two specs decide whether local AI runs well: RAM capacity, which sets the largest model you can load, and SSD capacity and speed, which set how many models you can keep and how fast they load. For most people, 32GB of DDR5 paired with a 1TB PCIe Gen4 NVMe SSD is the practical entry point. Coding assistants, AI agents and MoE models make 64GB of DDR5 and 2TB+ of NVMe the sweet spot.
- Why RAM Is the Core Bottleneck
- Estimating RAM Needs for Any Model
- Don't Forget the Context Window
- MoE Models Changed the Math
- RAM Sizing by Use Case
- Why Local AI Demands Serious Storage
- Typical Model File Sizes
- SSD Capacity & Speed
- Combined RAM + SSD Configurations
- Recommended Silicon Power Builds
- FAQ
Why RAM Is the Core Bottleneck
A local large language model (LLM) has to be held in memory while it runs. On a GPU, the weights sit in VRAM. On the CPU, or when a model is split between CPU and GPU, part or all of it sits in system RAM.
When there is not enough memory, the model either refuses to load or falls back to paging to disk, and responses slow to a crawl. That is why memory capacity, not CPU speed, is usually the first upgrade for local AI.
Estimating RAM Needs for Any Model
Start with the size of the model weights, then add memory for the context window and the rest of your system.
Example: a 12B model at Q4 → 12 × 0.6 ≈ 7–8 GB of weights. Plan for at least 1.5× the file size once you add context, the operating system and your other apps.
Don't Forget the Context Window (KV Cache)
While a model works, it keeps a KV cache of everything in the current conversation or document. The longer the context, the bigger the cache, and this is the most common sizing mistake.
- Short chats (4K–8K tokens) add relatively little memory.
- AI coding tools and agents typically need 64K tokens or more. At that length, the cache can be as large as the model itself. An 8B reasoning model with a ~5GB file can need around 10GB of extra cache at 64K.
- Newer models with more efficient attention need far less cache, so model choice matters as much as model size.
Learn how to size memory for long context, and how an 8-bit KV cache halves its footprint, in How Context Length and KV Cache Affect RAM for Local AI.
MoE Models Changed the Math
Many of the strongest open models of 2026 use a Mixture-of-Experts (MoE) design. A model labeled "35B-A3B" has 35 billion parameters in total, but only about 3 billion are active for each token.
- Memory follows the total size. A 30B-class MoE model at Q4 is roughly 20GB, the same as any 30B model.
- Speed follows the active size. Because only a few experts run per token, MoE models stay responsive even when the expert weights sit in system RAM instead of VRAM.
With llama.cpp-based tools, you can keep the attention layers on a mid-range GPU and place the expert weights in DDR5. A PC with an 8–16GB graphics card and 64GB of DDR5 can now run models that once needed a workstation GPU. See Run Bigger AI Models on a Smaller GPU with 64GB DDR5.
RAM Sizing by Use Case
| Use Case | Typical Models (Q4) | Minimum RAM | Recommended RAM |
|---|---|---|---|
| First steps / light chat | 4B–9B models | 16 GB | 32 GB |
| Everyday assistant | 12B-class models | 16 GB | 32 GB |
| AI coding assistant (64K+ context) | 27B dense or 30B–35B MoE | 32 GB | 64 GB |
| AI agent + RAG | 12B–35B plus an embedding model | 32 GB | 32–64 GB |
| AI image generation | Diffusion models plus LoRAs | 32 GB | 64 GB |
| Large MoE models / multiple models | 100B-class MoE with CPU offload | 64 GB | 128 GB |
Always install memory in matched pairs so it runs in dual-channel mode. Token generation speed depends heavily on memory bandwidth, and dual-channel DDR5-6000 delivers roughly twice the bandwidth of a single module. For more on speed versus capacity, see A Complete Guide to Memory Performance & Capacity.
Why Local AI Also Demands Serious Storage
Model libraries grow faster than most people expect.
- You download several models to compare quality and speed, often in more than one quantization.
- Coding, vision, embedding and reasoning models are separate downloads.
- LoRA adapters, diffusion checkpoints and RAG document sets keep piling up.
- New model versions arrive every few weeks, and most people keep the old ones "just in case."
An active setup can pass 500GB within a few months, and 1TB fills up quickly.
Typical Model File Sizes
| Model Type | Examples (Nov 2026) | Typical Download |
|---|---|---|
| Small chat models | Qwen 3.5 4B / 9B | 3–8 GB |
| Mid-size chat models | Gemma 4 12B | ~8 GB |
| Large dense models | Qwen 3.8 27B, Gemma 4 31B | 18–20 GB |
| Mid-size MoE models | Gemma 4 26B-A4B, Qwen 3.6 35B-A3B | 16–24 GB |
| Large MoE models | gpt-oss 120B, Qwen 3.5 122B | 65–81 GB |
| Image generation | Diffusion checkpoints | 5–20 GB each |
| LoRA adapters | Fine-tunes | MBs to a few GB each |
| RAG documents / vector DBs | Your own files | Tens of GB |
Model names and sizes change quickly. Always check the download size listed in Ollama or LM Studio.
SSD Capacity & Speed Recommendations
| Profile | Capacity | Suggested Silicon Power SSD |
|---|---|---|
| AI Chat | 1 TB | UD90 1TB (Gen4, up to 5,000 MB/s) |
| AI Creator | 2 TB | US75 2TB (Gen4, up to 7,000 MB/s) |
| AI Developer | 4 TB | US75 4TB with heatsink or UD90 4TB |
| AI Lab / Multi-Model | 2 TB + 4 TB | US85 2TB (Gen5, up to 10,300 MB/s) for active models + a 4TB Gen4 library drive |
Combined RAM + SSD Configurations
| User Profile | RAM | SSD |
|---|---|---|
| First-time local AI | 16–32 GB | 1 TB |
| Everyday chat / small assistant | 32 GB | 1 TB |
| Coding assistant / AI agent / RAG | 64 GB | 2–4 TB |
| Content creation / image generation | 64 GB | 2 TB |
| Large MoE models / power-user lab | 128 GB | 4 TB+ |
Recommended Silicon Power Builds
Silicon Power DDR5 and NVMe SSDs cover every tier above. Here are the combinations we recommend.
AI Starter
- Storm RGB DDR5-6000 CL36 32GB (2×16GB)
- UD90 1TB or US75 1TB
- Best for: local LLM beginners, 4B–12B chat models, small RAG projects
AI Creator
- Zenith RGB DDR5-6000 CL30 64GB (2×32GB)
- US75 2TB
- Best for: coding assistants, 27B–35B models, image generation, creators
Zenith RGB DDR5-6000 CL30 64GB (2×32GB)
Our top pick for AI Creator builds. Dual-channel 64GB in two slots, with AMD EXPO and Intel XMP 3.0 profiles.
Value Gaming DDR5-6000 CL30 64GB (2×32GB)
The same 6000 MT/s CL30 performance without RGB. Buy two kits for a 128GB AI Lab build.
US75 NVMe SSD
A high-bandwidth Gen4 drive that keeps model load and switch times short.
US85 NVMe SSD
Gen5 speed for frequent model switching and SSD expert streaming.
- Size RAM for the model plus its context window, not just the file size.
- 32GB is the practical entry point. 64GB is the sweet spot for coding, agents and MoE models.
- MoE models let system RAM do work that used to need a big GPU.
- Plan for at least 1TB of PCIe Gen4 NVMe storage, and 2–4TB if you experiment a lot.
Frequently Asked Questions
How much RAM and SSD do I need to start with local AI?
32GB of DDR5 paired with a 1TB PCIe Gen4 NVMe SSD is the practical entry point. It runs 4B–12B models comfortably and leaves room for your browser and other apps. 16GB works for small models only.
Why does local AI need so much SSD space?
Because a model file is rarely the only one you keep. Most users compare several models and quantizations, add coding, vision and embedding models, and collect LoRA adapters and RAG documents. An active setup often passes 500GB within a few months.
Is 64GB of RAM overkill for local AI?
Not anymore. AI coding tools need long context windows, and Mixture-of-Experts (MoE) models rely on system RAM when they do not fit in VRAM. For these workloads, 64GB (2×32GB) is the most practical upgrade.
Does memory speed matter, or only capacity?
Capacity decides whether a model loads. Speed and dual-channel operation decide how fast it answers whenever any part of the model runs from system RAM. Get the capacity you need first, then the fastest kit your platform supports.