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.
There is no single RAM or VRAM requirement for running an AI model locally. The amount you need depends on the exact model, its quantization, the length of the context you want to use, and the software running it. As a practical starting point, NVIDIA’s 2026 guidance pairs a 4B model with 6–8 GB of GPU memory, 9B–12B models with 12–16 GB, and a 27B model with 24 GB or more. Those are vendor starting points, not guarantees for every model or setup.
To size a system responsibly, check the target model and runtime first, then leave room beyond the model weights for the KV cache and other software. Also distinguish discrete GPU VRAM from system RAM and shared unified memory: they are different pools, and one cannot be treated as a universal substitute for another.
As an Amazon Associate I earn from qualifying purchases.
Free tools Windows power users keep installed
One-click scans. No signup required.
What do RAM and VRAM do when you run a model locally?
VRAM is memory physically associated with a discrete graphics card. When inference runs on that GPU, the model’s weights and parts of its working state need available GPU memory. If the model does not fit entirely in VRAM, some runtimes can place part of the work in system memory, but whether that is supported and how it affects performance depends on the backend and configuration.
System RAM is the main pool for CPU-only inference and also supports the operating system, model files, and other running applications. A computer with a discrete GPU has both system RAM and VRAM; adding more of one does not automatically increase the capacity of the other. A card with 8 GB of VRAM does not become a 16 GB card because the computer has 16 GB of system RAM. Likewise, installing a larger GPU does not provide extra host RAM for a CPU-only runtime.
#1 Best Overall
- Read Before You Buy — No Video Output: These adapters support charging and USB 2.0 data transfer, but cannot transmit video signals. Except for standard USB webcams (which use USB data only), they are not compatible with HDMI/DisplayPort cables, video-capable USB-C hubs, or docking stations with video output.
- Convert USB-A Ports to USB-C: Designed to connect USB-C earphones, cables, flash drives, card readers, and other USB-C accessories to standard USB-A ports. Plug-and-play with no drivers or software required.
- Aluminum Alloy Housing: Built with a sturdy aluminum alloy shell that aids in heat dissipation and protects against daily wear and scratches. Designed to maintain a stable and secure connection.
- Compact & Travel-Friendly: The ultra-compact design allows the adapter to stay plugged into your device without blocking adjacent ports or adding bulk, reducing wear and tear on your original USB ports.
- 12-Month Warranty: Backed by a 12-month manufacturer warranty for peace of mind. Designed to meet strict quality control standards for reliable everyday performance.
Apple Silicon and other unified-memory systems are a different case: CPU and GPU functions share a memory pool. The number printed in the product specification is therefore not all available to the model. The operating system, display, runtime, and active applications consume part of it. Compare the usable unified memory after those reservations with the model and its runtime needs.
Memory capacity determines which models can be loaded and run in a given configuration. Capacity alone does not predict speed: memory bandwidth, the inference backend, the amount of work assigned to the GPU, and the model itself also affect the experience. The figures below are memory-fit starting points, not performance benchmarks.
How much VRAM do common model sizes need?
Use model-size tiers as an initial filter, not as a shopping guarantee. NVIDIA’s current local-LLM playbook lists these example GPU-memory pairings:
| Available memory | NVIDIA example starting point | How to interpret it |
|---|---|---|
| 6–8 GB GPU VRAM | Qwen 3.5 4B | A small-model starting tier; confirm the selected quantization, context, and runtime. |
| 12–16 GB GPU VRAM | Qwen 3.5 9B or Gemma 4 12B | A midrange example tier, not a promise that every 9B or 12B model will fit. |
| 24 GB or more GPU VRAM | Qwen 3.6 27B | A larger-model example; memory use still varies with quantization and context length. |
| 64 GB or 128 GB unified memory | DGX Spark platform | A documented platform example, not a general minimum for local inference. |
These recommendations are specifically NVIDIA’s suggested starting points for the named models and hardware tiers. Do not assume that every model with a similar parameter count will have identical memory requirements. Architecture, quantization, context, and runtime all matter. A sparse or mixture-of-experts model may have a large total set of weights but activate only part of them for each token, while still requiring storage for the complete model files. That makes a simple parameter-count rule especially unreliable.
A separate reference point comes from the Yale Center for Research Computing, which gives approximate VRAM estimates for unquantized inference: 7B at 10+ GB, 13B at 20+ GB, 30B at 40+ GB, 70B at 80+ GB, and 305B at 400+ GB. These are approximate unquantized figures, not direct equivalents to NVIDIA’s quantized or model-specific examples. Yale advises checking the individual model’s page on Hugging Face or Ollama for its requirements.
Do not combine the two tables as if they were one conversion chart. A 7B unquantized estimate and a vendor’s quantized 9B recommendation describe different precision, software, and workload assumptions. Use each figure only within the context in which it was published.
Rank #2
- 5-in-1 USB-C Hub: Experience comprehensive connectivity featuring a Power Delivery input, two USB-A 2.0 ports, a USB-A 3.0 port, and an HDMI port. (Note: The USB-C power delivery input port is only for connecting an external wall charger to power your laptop and cannot power peripheral devices.)
- 90W Pass-Through Charging: Achieve optimal charging with 90W pass-through power to your laptop, supported by a total input of 100W, with the hub reserving 10W for operational efficiency. (Note: Wall charger not included.)
- Quick Data Transfers: Accelerate your productivity with rapid data transfers using a high-speed 5Gbps USB 3.0 port and two 480Mbps USB 2.0 ports.
- 4K HDMI Display: Enhance your visual experience with a hub capable of delivering 4K resolution at 30Hz in both mirror and extend modes. Please note that this hub is compatible with MacBook (macOS 12 and newer), Windows 10 and 11, ChromeOS, and laptops equipped with DP Alt Mode and Power Delivery. Note: This device is not compatible with Linux.
- What You Get: Anker USB-C Hub (5-in-1, 4K HDMI), welcome guide, 18-month warranty, and our friendly customer service.
Why can a model need more memory than its weights?
Model weights are only part of the allocation. During inference, a runtime also uses memory for the KV cache, which stores information needed as the model processes context. Longer context and concurrent requests can increase cache needs, so a model that loads at one context length may require more memory at a longer one or when serving multiple sessions. Chat history, retrieved documents, and generated output all contribute to the amount of context being maintained.
NVIDIA’s llama.cpp documentation makes that distinction concrete: its particular Qwen3.6-35B-A3B MTP example estimates about 30 GB of free memory to account for the model and cache needs. That is an estimate for that named example, not a universal 35B requirement. The same documentation, last updated October 9, 2026, lists DGX Spark with 64 GB or 128 GB of unified memory as supported hardware. Those configurations do not establish a minimum for other computers or runtimes.
Usable memory can also be lower than the capacity printed on a GPU or computer. The operating system, display, runtime, driver allocations, model-loading buffers, and other applications may reserve some of it. A browser with many tabs, a game, a desktop compositor, or another model can leave less headroom than a benchmark assumes. When assessing a fit, compare the model’s expected use with memory available to the inference backend, not just the system’s headline capacity.
A model that barely loads can still be an impractical choice. Near the limit, changing the context window, opening another application, or starting a second request may trigger an out-of-memory error. Leave a margin instead of treating every advertised gigabyte as allocatable model space.
How does quantization change the memory requirement?
Quantization stores model weights at lower numerical precision, which can reduce their memory footprint. It often involves a trade-off in quality or performance, and the result depends on the model, quantization method, and runtime. A model’s parameter count by itself therefore does not tell you how much memory a particular downloadable version will use.
Puget Systems provides a configuration-specific example from its Llama 3.1 8B tests. In a Transformers setup, the native BF16 version used just over 15 GB of VRAM. In its llama.cpp tests, the 8-bit and 4-bit versions required about half and about one quarter of the native-precision VRAM, respectively. Those ratios describe the tested setup; they should not be applied as exact conversion rules to every model or backend.
Rank #3
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
When comparing model files, identify the specific quantized variant and the runtime that supports it. “4-bit” is not a complete specification by itself: implementations can use different formats and overheads. A lower-bit version may fit in less memory, but the amount saved and the effect on output quality or speed are not established by parameter count alone. Confirm that the downloaded file is compatible with the loader you intend to use, rather than assuming any file with the same model name can be substituted.
How do you estimate the memory for your own setup?
Start with the model and workload you actually plan to use, rather than buying hardware based on a generic “B” size rule. These checks narrow the answer to a useful estimate:
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstall- Choose the exact model. Record its name, release, and variant, not only its parameter count. Distinguish dense models from sparse or mixture-of-experts architectures, and note whether the download includes adapters or other components.
- Choose the quantization and runtime. Memory use can differ substantially between precision levels and inference backends. Confirm that the runtime supports the model variant you intend to load and determine whether it keeps all layers on the GPU or allows partial offload.
- Set a realistic context length and request load. Longer context and concurrent requests can increase KV-cache requirements. If you expect several sessions, batch requests, or a retrieval workflow that adds large documents, size for that use rather than a single short prompt.
- Identify the memory pool used by inference. For a discrete-GPU workflow, check available VRAM after display reservations. For CPU inference, check available system RAM after the operating system and applications. For a unified-memory system, assess the shared pool available to both CPU and GPU functions.
- Check the model’s own memory guidance. Look at the exact model variant’s current Hugging Face or Ollama page and the runtime documentation. Use a source that matches the model, precision, context assumptions, and backend instead of transferring a number from a different setup.
- Leave headroom. Do not target a configuration where the estimated model allocation consumes every available gigabyte. The cache, runtime, display, operating system, and other active programs can all affect whether inference fits.
- Test the intended workload. A successful model load is only the first check. Exercise the planned context length and number of sessions, then observe memory use and stability before treating the hardware as adequate.
This process produces a configuration-specific estimate rather than a universal RAM multiplier. There is no authoritative host-RAM formula that applies across runtimes, model architectures, and operating systems, so avoid treating system RAM as a fixed multiple of VRAM or model size.
What if you are comparing hardware?
Compare usable memory capacity first, but do not stop there. Memory bandwidth affects how quickly weights and cache data can be moved, while backend support determines whether the hardware can actually execute the chosen model efficiently. Power requirements, cooling, physical size, and the number of other tasks the machine must handle also matter.
A 24 GB graphics card is a sensible product category to investigate for workloads in NVIDIA’s 24 GB-and-up example tier, including its Qwen 3.6 27B starting point. It is not a guarantee that every 27B model, context length, or quantization will fit. Before buying, verify the card’s genuinely usable VRAM, the backend’s support for its architecture, the expected model file, and room for cache. Compare the complete purchase cost and operating constraints rather than selecting solely by the largest number on the box; no universal price comparison follows from the memory tiers alone.
High-capacity unified-memory computers are another category for evaluation. NVIDIA’s 64 GB and 128 GB DGX Spark entries document a particular platform, not an equivalence claim for every consumer computer with a similar memory number. Check how much memory is shared with the operating system, which runtimes support the device, and whether the workload needs discrete-GPU features.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Rank #4
- Dual Converters, Infinite Potential：Includes 2× USB C male to USB A female adapters and 2× USB A male to USB C female adapters. Perfect for a wide range of uses—tablets with Bluetooth keyboards, expand USB ports on macbook, and more. Two different converters for all your daily needs
- Next-Level 10Gbps & 3A Charging: No more slow 480Mbps, this usb to usb c adapter has a transfer speed of up to 10Gbps, allowing you to do more transferring in less time. This usb adapter fits both USB A and USB C charger, supporting up to 3A fast charging
- Upgraded Exquisite Craftsmanship: With an aluminum alloy housing and metal connector, the usbc to usb adapter is extremely durable and sturdy. Rigorously tested to withstand more than 10,000 times of plugging and unplugging, ensuring long-lasting performance
- Broad Compatible: The usb c to usb adapter widely supports all USB C/ USB A devices like laptops, tablets, cellphones, car chargers, and phone chargers. Such as compatible with MacBook Pro/Air 2023/2022, Thunderbolt 4/3 Devices,Apple MagSafe Watch 9/8/7/SE/Ultra, iPad Pro 2022/2021, Samsung Galaxy S23/S20/S10, and iPhone 17/16/15 Pro. Plug and play
- Please Note: To reach 10Gbps speed, keep the cable under 3.3 ft. For USB A Male to USB C adapters, try flipping the USB C connector. USB C Male to USB A adapters support bidirectional 10Gbps transfer within 3.3 ft
What should you do if a model does not fit in VRAM?
First check whether the selected quantization, context length, or concurrent workload can be reduced while preserving the use case. A shorter context or fewer simultaneous sessions may reduce KV-cache pressure, while a supported lower-precision model may reduce weight memory. These changes can affect answer quality, throughput, or features, so validate them with representative prompts.
If the runtime supports offloading, it may be able to place some work in host memory instead of keeping everything in GPU memory. The availability and performance impact depend on the backend. Transfers between system RAM and VRAM can make generation substantially slower, and an offloaded configuration may behave differently from a fully GPU-resident one. A successful load therefore does not mean the model will run at the speed you expect.
If neither adjustment is acceptable, choose a larger memory pool or a smaller model. Remember that adding system RAM does not enlarge a discrete card’s VRAM, and adding VRAM does not help a CPU-only program unless the backend can use that GPU. Unified-memory systems avoid a hard split but still have one shared capacity limit.
NVIDIA summarizes the practical constraint as “Memory determines what you can run.” Its current recommendations are helpful for shortlisting, but a final fit decision belongs to the exact model, quantization, context, and runtime you plan to use.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Frequently Asked Questions
Does a 7B model always need 10 GB of VRAM?
No. Yale Center for Research Computing gives 10+ GB as an approximate estimate for unquantized 7B inference. A quantized version or a different runtime can have a different footprint, so treat that figure as a reference point rather than a fixed requirement.
Can I use system RAM instead of VRAM for local AI?
Some inference setups can offload work to host memory when a model does not fit fully in GPU memory, but support and performance vary by backend. CPU-only inference uses system RAM as its main model memory pool; it is not the same configuration as GPU inference.
Best Value
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
Where should I look up the requirements for a specific model?
Check the exact model variant’s current page on Hugging Face or Ollama, then verify the inference runtime’s documentation for that model and quantization. Yale Center for Research Computing specifically recommends checking those model pages because generic parameter-count estimates are approximate.
Does a longer context window change the GPU requirement?
Yes. Longer context increases KV-cache use, and concurrent requests can increase it further. A model that fits with a short conversation may need substantially more free memory when processing long documents or several sessions.
Is 24 GB of VRAM enough for any 27B model?
No. NVIDIA uses 24 GB or more as a starting tier for its Qwen 3.6 27B example. The exact quantization, context length, cache, runtime overhead, and architecture determine whether another 27B model fits.
Is unified memory the same as having separate RAM and VRAM?
No. Unified-memory systems share one pool between CPU and GPU functions, while a discrete-GPU computer has separate system RAM and VRAM pools. Evaluate the memory available after system reservations and confirm that the runtime supports the platform.
What is the safest way to avoid an out-of-memory error?
Use the exact model and runtime documentation, size for the intended context and concurrency, and leave headroom for the KV cache, operating system, display, and other applications. Test the real workload rather than checking only whether the model file can be opened.
The Bottom Line
Bottom line: Match memory to the exact model, precision, context length, concurrency, and backend. As NVIDIA starting points, think 6–8 GB for its 4B example, 12–16 GB for its 9B–12B examples, and 24 GB or more for its 27B example. Treat those as filters, not guarantees: reserve space for KV cache and software, distinguish VRAM from system RAM or unified memory, and verify the model’s current requirements before buying hardware.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsQuick 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.
