What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
For a practical starting point, plan on 16GB of system RAM for local 7B–8B models, 16–32GB for many 13B–14B models, and 64GB or more for 70B-class models. Those are planning ranges, not guarantees. A model’s quantization, context length, runtime, and the division of work between CPU and GPU all affect what will fit and how quickly it runs. System RAM and GPU VRAM are separate memory pools, so check both before choosing a model or upgrading a computer.
How much RAM does a local AI model need by size?
Use parameter count as your first estimate, then adjust for quantization, context, and hardware placement. The table combines guidance from NIH Biowulf, Ollama’s Llama 2 documentation, NVIDIA, and LM Studio; it is a planning guide, not a promise that a particular model will load or run well.
As an Amazon Associate I earn from qualifying purchases.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11| Model size | Typical Q4 memory signal | Practical system RAM target | What to expect |
|---|---|---|---|
| 1–3B | About 2GB of VRAM in NIH Biowulf’s table | 8GB can work | Suitable for lightweight assistants, but the operating system, applications, and context still use memory. |
| 7–8B | About 6–8GB of VRAM in NIH Biowulf’s table | 16GB recommended | Ollama lists at least 8GB of RAM for a 7B model; 16GB gives a desktop more room for other work. |
| 13–14B | About 12–16GB of VRAM in NIH Biowulf’s table | 16–32GB | Ollama lists at least 16GB of RAM for a 13B model. More system memory helps when context or CPU offload adds demand. |
| 30–35B | Roughly 20GB-plus of Q4 VRAM, depending on runtime | 32GB or more | NVIDIA’s examples put this class in the 24GB-or-more GPU tier, but the exact fit varies by model. |
| 70B+ | About 35–40GB of Q4 VRAM in NIH Biowulf’s table | 64GB or more | Ollama’s general guidance is at least 64GB of RAM. Multi-GPU systems or unified-memory hardware are common approaches. |
These columns describe different resources: the Q4 signal is a GPU-memory estimate, while the target is system RAM. Neither number includes every model, application, and runtime overhead in every configuration. A model may load with less headroom than recommended and still be unpleasantly slow or leave too little memory for other tasks.
How much RAM do I need for a 7B model?
For a 7B model, 8GB of system RAM is the stated minimum in Ollama’s Llama 2 documentation, but 16GB is a more comfortable desktop target. If you intend to run a long context, keep a browser or development tools open, or rely on CPU memory because the GPU is short on VRAM, allow more room.
#1 Best Overall
- Capacity: 16GB(2 x 8GB)
- Tested Frequency Profile 1: PC5-48000 (6000MT/s)
- Tested Timings: 36-46-46-110
- Feature Overclock: XMP 3.0 & EXPO overclocking supported
- On-Die ECC
Why 8GB is a floor rather than a comfort target
The model is not the only process using memory. The operating system and local runtime need space, and other open applications compete for the same system RAM. A model that uses most of an 8GB machine’s available memory can leave little capacity for chat history, document retrieval, or ordinary desktop use.
When 16GB makes sense
Choose 16GB if you want a 7B–8B model to coexist with normal desktop applications, or if you do not yet know how much context your workflow will use. NIH Biowulf’s table indicates about 6–8GB of VRAM for a typical Q4 model in this size range; that is a separate GPU-memory consideration, not a requirement that can be satisfied by adding system RAM.
Can I run a 13B model on 16GB RAM?
Often, 16GB is a reasonable starting point for a 13B model, but it is not a guarantee of a comfortable setup. Ollama’s Llama 2 documentation gives 16GB as the general RAM figure for 13B models; 32GB gives more margin for longer context, applications, and CPU offload.
The difference between “can run” and “runs well” matters. A system may be able to load a quantized model but respond slowly if too much work falls to the CPU, or become constrained when a large prompt and chat history are added. Check the model’s actual quantization and runtime requirements rather than treating the 13B label as a complete specification.
How much VRAM does a 70B model need?
NIH Biowulf’s table gives about 35–40GB of VRAM for a 70B-class model at Q4, while the practical system-RAM target is at least 64GB. A single GPU may not have enough VRAM for full GPU placement, so multi-GPU setups, Apple-style unified memory, or partial CPU offload can be relevant, each with its own performance trade-offs.
Ollama’s general guidance also says 70B models need at least 64GB of RAM. NVIDIA NIM, using a different rough sizing approach, lists Llama 70B at about 131GB and advises reserving 5–10GB for the operating system and other processes. These figures should not be treated as equivalent measurements: the Q4 table and NIM sizing do not establish the same model variant, runtime, or placement assumptions. Use the requirements for the exact model and software you plan to run.
What is the difference between RAM and VRAM for local AI?
System RAM is the computer’s main memory; VRAM is memory on a graphics card. They are distinct pools, and extra system RAM does not turn into extra GPU VRAM. NIH Biowulf explicitly cautions that GPU memory means VRAM, not a scheduler’s --mem value.
Rank #2
- Boosts System Performance: 32GB DDR5 RAM laptop memory kit (2x16GB) that operates at 5600MHz, 5200MHz, or 4800MHz to improve multitasking and system responsiveness for smoother performance
- Accelerated gaming performance: Every millisecond gained in fast-paced gameplay counts—power through heavy workloads and benefit from versatile downclocking and higher frame rates
- Optimized DDR5 compatibility: Best for 12th Gen Intel Core and AMD Ryzen 7000 Series processors — Intel XMP 3.0 and AMD EXPO also supported on the same RAM module
- Trusted Micron Quality: Backed by 42 years of memory expertise, this DDR5 RAM is rigorously tested at both component and module levels, ensuring top performance and reliability
- ECC Type = Non-ECC, Form Factor = SODIMM, Pin Count = 262-Pin, PC Speed = PC5-44800, Voltage = 1.1V, Rank And Configuration = 1Rx8
If a model is placed entirely on a GPU, its weights and other runtime needs must fit in available VRAM. When VRAM is insufficient, some software can place part of the model in system RAM and run part on the CPU. That can make a larger model possible, but NIH Biowulf warns that Ollama’s CPU fallback is very slow except for the smallest models. Partial offload is therefore a compatibility option, not a substitute for adequate GPU memory when speed matters.
How much does quantization change memory needs?
Quantization stores model weights with fewer bits, reducing the memory needed to load them. NIH Biowulf says 4-bit quantization reduces memory to about 25% of the original; Ollama’s model documentation uses q4 for a 4-bit variant and notes that higher-bit variants use more memory.
Q4 can be the practical way to fit a model on consumer hardware, but it is not a free reduction. NVIDIA cautions that aggressive quantization can reduce response quality. The right choice depends on whether fitting the model is the priority or whether preserving more of the original model’s quality is worth using more memory.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitches| Variant | Memory implication | Decision point |
|---|---|---|
| Q4 / 4-bit | NIH Biowulf estimates about one-quarter of original memory for 4-bit quantization. | Useful when memory is the binding constraint; quality can be affected by aggressive quantization. |
| Q5 or Q8 | Higher-bit variants use more memory than Q4, according to Ollama’s model documentation. | Check the exact variant’s size and fit before selecting it; the available data does not establish a universal memory figure for these levels. |
| FP16 | More memory-intensive than low-bit quantized variants; no universal amount is established here. | Use the specific model and runtime’s stated requirements rather than extrapolating a guaranteed figure. |
Why does context length add to memory use?
Context is the information the model can consider at one time: the prompt, conversation history, tool outputs, and retrieved documents. NVIDIA explains that longer context uses more memory, so a model’s weight-file size alone cannot tell you whether a long-document workflow will fit.
A short exchange and a workflow that feeds a large document collection into a 128K-token context are different workloads, even if they use the same model file. Leave more headroom when you expect long conversations, tool use, or retrieval-augmented prompts. The available guidance establishes the direction of the effect, not one fixed number of additional gigabytes per token count.
Is 16GB, 32GB, or 64GB system RAM the right tier?
Choose based on the model class and what else the computer must do. More system RAM helps with CPU placement, multitasking, and room for context, but it does not solve a GPU VRAM shortage when you need full GPU offload.
| System RAM tier | Best fit in this guide | Pairing and trade-off |
|---|---|---|
| 16GB | Practical target for 7B–8B; starting point for many 13B models | For a 7B–8B Q4 GPU workload, NIH’s signal is about 6–8GB VRAM. More RAM does not guarantee a longer context or faster generation. |
| 32GB | More headroom for 13B–14B, longer context, multitasking, or partial CPU offload | NIH’s Q4 signal for 13B–14B is about 12–16GB VRAM. A 24GB-or-more GPU is the tier NVIDIA’s examples associate with 30–35B models. |
| 64GB+ | Practical system-memory target for 70B-class models | NIH’s Q4 estimate for 70B is about 35–40GB VRAM. Multi-GPU or unified memory may be needed for larger models; actual needs depend on variant and runtime. |
These are planning pairings, not component compatibility guarantees. Verify the model variant, software support, memory type, and actual available capacity on the machine you intend to use.
Rank #3
- Disclaimer: Maximum Speed requires overclocking/PC BIOS adjustments. Maximum speed and performance depend on system components, including motherboard and CPU
- AMD EXPO & Intel XMP 3.0 Compatible Only: Dual memory profiles allow you to easily select optimized settings for your platform, whether you’re running an AMD or Intel processor
- Dynamic RGB Lighting: Individually addressable RGB lighting delivers vibrant effects through a sleek, understated panoramic diffuser
- Onboard Voltage Regulation: Onboard voltage regulation for reliable power at high frequencies
- Maximum Bandwidth and Tight Response Times: Optimized for peak performance on the latest AMD and Intel DDR5 motherboards
Will 4-bit quantization fit on my laptop?
Possibly, but “4-bit” alone cannot answer that. Compare the exact model’s memory requirements with the laptop’s available system RAM and GPU VRAM or unified memory, then account for context and runtime overhead. NIH’s Q4 figures provide useful signals by model class, but they are not guarantees for every laptop or implementation.
LM Studio says a dedicated GPU is optional and recommends at least 4GB of dedicated VRAM for Windows. That recommendation is not the same as a promise that a particular Q4 model will run quickly or fit fully on the GPU. If the laptop must offload work to CPU and RAM, expect a speed penalty; if the model does not fit comfortably, select a smaller model or lower-memory variant rather than assuming that the quantization label solves every constraint.
Can I run a local LLM without a GPU?
Yes, software such as Ollama can fall back to CPU, and LM Studio says a dedicated GPU is optional. The trade-off is speed: NIH Biowulf describes Ollama’s CPU fallback as very slow except for the smallest models.
A CPU-only machine is more plausible for lightweight 1–3B models than for large models intended for GPU inference. System RAM can make CPU execution possible, but it does not make CPU inference equivalent to having enough VRAM for GPU placement. For a machine without a dedicated GPU, start with a small model and treat larger sizes as experiments whose responsiveness depends on the particular hardware and runtime.
Recommended Free Tools
How much overhead should I leave for the operating system and runtime?
Do not plan to give every available gigabyte to model weights. NVIDIA NIM’s rough guidance reserves 5–10GB for the operating system and other processes, in addition to its model-sizing examples. Other applications, the local runtime, and context also compete for memory.
NVIDIA NIM notes an additional 16GB of Docker shared memory in some multi-GPU non-NVLink deployments. This is a configuration-specific requirement, not a general allowance every local-model user must add. Follow the deployment documentation for the runtime and hardware you actually use.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How do I choose a model that fits my computer?
Start from the available memory and workload rather than choosing by parameter count alone. NVIDIA’s general recommendation is to use the most powerful model that fits comfortably in the GPU’s memory.
Rank #4
- Disclaimer: Maximum Speed requires overclocking/PC BIOS adjustments. Maximum speed and performance depend on system components, including motherboard and CPU
- Intel XMP 3.0 Compatible Only: Customize and save your own XMP profiles via iCUE to tailor performance by app or task for greater efficiency
- Dynamic Ten-Zone RGB Lighting: Illuminate your system with ten individually addressable, ultra-bright RGB LEDs per module, encased in a panoramic light bar for vivid RGB lighting from any viewing angle
- Onboard Voltage Regulation: Enables easier, more finely-tuned, and more stable overclocking through CORSAIR iCUE software than previous generation motherboard control
- Maximum Bandwidth and Tight Response Times: Optimized for peak performance on the latest Intel DDR5 motherboards
- Identify the memory pools. Check installed system RAM and, if present, the GPU’s dedicated VRAM. On hardware using unified memory, check the total available pool and the software’s requirements.
- Choose a model-size range. Use the sizing table to shortlist 7B–8B, 13B–14B, or larger models based on your system-memory target and available GPU resources.
- Check the precise quantization. Find the exact variant, such as Q4, Q5, Q8, or FP16. Higher-bit variants use more memory; Q4 reduces the requirement substantially but may affect quality.
- Account for the actual context. Include the prompt, chat history, tool outputs, and documents you expect to use. A short conversation is not a reliable proxy for a long-context workload.
- Decide whether CPU offload is acceptable. If VRAM is short, partial CPU placement may enable loading, but it can sharply reduce speed. CPU-only use is also possible, with the same performance caveat for all but the smallest models.
- Leave operating headroom. Reserve memory for the OS, runtime, and other open applications; NVIDIA NIM’s rough estimate is 5–10GB for OS and other processes.
- Verify the software’s current requirements. Model catalogs, supported variants, and runtime needs can change. Check the documentation for the exact model and version before downloading or upgrading hardware.
What should I upgrade first: RAM, GPU memory, or storage?
Upgrade the resource that is actually constraining your intended workload. If you are below the system-RAM target for your model class or multitasking needs, more RAM can help. If you want full GPU placement but the model’s VRAM estimate exceeds the GPU’s capacity, additional system RAM will not remove that VRAM limit.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Storage is needed for model files, but the available sizing guidance does not establish a universal disk-space figure for each parameter class or quantization. Check the download size of the exact model variant and leave capacity for the files and applications you intend to keep. For a RAM upgrade, confirm DDR generation, laptop SO-DIMM versus desktop DIMM, maximum supported capacity, and available slots before buying; memory type and configuration are system-specific.
Why do published memory estimates disagree?
They may describe different model variants, quantization levels, runtimes, context sizes, or hardware placement. For example, NIH Biowulf’s table gives a Q4 VRAM signal, while NVIDIA NIM lists about 15GB for Llama 8B and 131GB for Llama 70B under its rough sizing rule and separately allows 5–10GB for OS and other processes. Those numbers should not be averaged or presented as contradictory measurements of an identical setup.
Ollama’s figures are broad minimum RAM guidance for its Llama 2 documentation, dated 2023-07-18: at least 8GB for 7B, 16GB for 13B, and 64GB for 70B. Treat such minimums as product-specific guidance rather than universal hardware standards. When estimates differ, prioritize the requirements published for your exact model build, quantization, runtime, and deployment configuration.
Frequently Asked Questions
Does adding more RAM make a local AI model generate answers faster?
Not necessarily. Extra RAM can prevent memory pressure or make CPU offload possible, but it does not by itself increase GPU compute or GPU memory bandwidth; the effect depends on what was limiting the system.
Can I upgrade a laptop’s RAM for local AI?
Some laptops allow upgrades and others have memory soldered to the motherboard. Check the manufacturer’s specifications for maximum capacity, supported DDR generation, and whether the machine uses replaceable SO-DIMM modules before buying.
How much storage should I reserve for local AI models?
There is no universal storage amount in the cited sizing guidance: download size varies by model and variant. Check the exact model file size before downloading, and remember that storage capacity is separate from the RAM or VRAM needed to run it.
Does Apple unified memory count as both RAM and VRAM?
Unified-memory systems share a memory pool between the CPU and GPU, unlike a typical desktop with separate system RAM and graphics-card VRAM. Available capacity still has to cover the operating system, applications, and model workload, and software support affects which models can run.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
