Memory fits
YOUR OWN HARDWARE
NVIDIA GeForce RTX 4090
24 GB VRAM
PRACTICAL TECHNOLOGY YOU CAN RUN AND CONTROL
Understand the tech.
Choose what fits.
Local AI, self hosting, home servers, hardware and website hosting.
Useful tools and clear guides for deciding what to run, what it needs and what it costs.
EXPLORELocal AISelf hostingHome servers
Explore HostingLinks
Start with the problem you want to solveLOCAL AI
Run AI on your own hardware
Understand models, memory, runtimes, privacy and the trade-offs between local and hosted AI.
Explore local AISELF HOSTING
Run useful software yourself
Explore photos, files, documents, media, automation and other services you can host.
Explore self hostingHOME SERVERS
Build a useful home server
Start with the workload, then plan hardware, storage, power, backups and maintenance.
Explore home serversHARDWARE
Choose hardware for the job
Make sense of GPUs, RAM, mini PCs and storage without relying on generic shopping lists.
Explore hardwareWEBSITES & HOSTING
Put the right site in the right place
Compare hosting categories, ongoing costs and the maintenance you are taking on.
Explore hostingTOOLS
Use the calculators and finders
Check memory, costs, model files and project options with assumptions made visible.
Explore toolsDOMAINS
Research names, value and authority
Understand domain value, backlink metrics, history, marketplaces and sponsored-content economics.
Explore domainsWhat can my computer run?
START WITH WHAT YOU OWNEnter your memory to find local models that fit. Compare the space left over, then explore a model in the calculator.
New to local AI? Start here
First, find your memory: on Windows, open Task Manager → Performance and check dedicated GPU memory and system Memory. On an Apple Silicon Mac, check About This Mac. Step-by-step instructions for Windows, Mac and Linux.
Then choose your settings: select the runtime you plan to use. For an initial Ollama or LM Studio comparison, try 4,096 context tokens and Balanced weights. Longer context holds more conversation but usually needs more memory. Quantization makes model weights smaller, with a possible quality trade-off.
The numbers already filled in are examples, not a scan of your computer. Replace them with your own specifications. Understand every setting.
Use available memory in GiB (1 GiB = 1,024³ bytes). One conversation, full GPU loading. Mac results reserve memory for macOS; dedicated GPU results must also fit system RAM. Memory fit does not guarantee speed or runtime compatibility.
| Model | Format | VRAM target | Memory left | Explore |
|---|---|---|---|---|
| DeepSeek R1 Distill Qwen 1.5BDeepSeek · 16 GiB RAM target | Q4_K_M | 2.5 GiB | 13.5 GiB | |
| DeepSeek R1 Distill Qwen 7BDeepSeek · 16 GiB RAM target | Q4_K_M | 6.4 GiB | 9.6 GiB | |
| DeepSeek R1 Distill Qwen 14BDeepSeek · 24 GiB RAM target | Q4_K_M | 12.0 GiB | 4.0 GiB | |
| Gemma 3 1B InstructGoogle · 16 GiB RAM target | Q4_K_M | 2.0 GiB | 14.0 GiB | |
| Gemma 3 4B InstructGoogle · 16 GiB RAM target | Q4_K_M | 4.8 GiB | 11.2 GiB | |
| Gemma 3 12B InstructGoogle · 16 GiB RAM target | Q4_K_M | 12.0 GiB | 4.0 GiB | |
| Llama 3.1 8BMeta · 16 GiB RAM target | Q4_K_M | 7.2 GiB | 8.8 GiB | |
| Llama 3.2 1B InstructMeta · 16 GiB RAM target | Q4_K_M | 2.2 GiB | 13.8 GiB | |
| Llama 3.2 3B InstructMeta · 16 GiB RAM target | Q4_K_M | 4.1 GiB | 11.9 GiB | |
| Ministral 3 3B InstructMistral AI · 16 GiB RAM target | Q4_K_M | 5.1 GiB | 10.9 GiB | |
| Ministral 3 8B InstructMistral AI · 16 GiB RAM target | Q4_K_M | 8.5 GiB | 7.5 GiB | |
| Ministral 3 14B InstructMistral AI · 24 GiB RAM target | Q4_K_M | 11.9 GiB | 4.1 GiB | |
| Qwen 2.5 0.5B InstructQwen · 16 GiB RAM target | Q4_K_M | 1.5 GiB | 14.5 GiB | |
| Qwen 2.5 1.5B InstructQwen · 16 GiB RAM target | Q4_K_M | 2.3 GiB | 13.7 GiB | |
| Qwen 2.5 7B InstructQwen · 16 GiB RAM target | Q4_K_M | 6.4 GiB | 9.6 GiB | |
| Qwen 2.5 14B InstructQwen · 24 GiB RAM target | Q4_K_M | 12.0 GiB | 4.0 GiB | |
| Qwen3 0.6BQwen · 16 GiB RAM target | Q4_K_M | 2.5 GiB | 13.5 GiB | |
| Qwen3 1.7BQwen · 16 GiB RAM target | Q4_K_M | 3.1 GiB | 12.9 GiB | |
| Qwen3 4BQwen · 16 GiB RAM target | Q4_K_M | 4.9 GiB | 11.1 GiB | |
| Qwen3 8BQwen · 16 GiB RAM target | Q4_K_M | 7.5 GiB | 8.5 GiB | |
| Qwen3 14BQwen · 24 GiB RAM target | Q4_K_M | 11.8 GiB | 4.2 GiB |
Compare up to three models
All columns use the computer, runtime, context and weight format selected above. You can include models that do not fit to see the difference.
What could I run after an upgrade?
Compare a possible memory configuration with your computer above. Runtime, context and weight format stay the same. These are capacity scenarios, not speed predictions.
21 models now → 23 with this configuration. 2 additional models fit.
Current: 16 GiB VRAM and 32 GiB RAM. Proposed: 24 GiB VRAM and 32 GiB RAM.
| Model | Format | Memory target | System RAM target |
|---|---|---|---|
| Gemma 3 27B Instruct | Q4_K_M | 22.9 GiB | 24 GiB |
| gpt-oss 20B | MXFP4 | 16.4 GiB | 24 GiB |
Mac unified memory usually requires choosing a different machine. PC upgrades also depend on power supply, case space and runtime support. Capacity alone is not a purchase recommendation. Proposed configurations are not included in share links.
Explore hardware options in the model calculator · See measured speed reference results. Hardware matches use the model selected in the calculator, not this proposed configuration. Provider links there may be affiliate links.
Memory left is the difference between your capacity and our target, which already includes headroom. Gemma estimates cover text only. gpt-oss requires a backend that keeps MXFP4 weights compressed. Looking for ChatGPT, Claude or Gemini?
01 Configure your model
Curated models / hardware examples
A compact starting point for local chat, writing, and everyday tasks.
4.85 effective bits per weight · 4.85 GB estimated weights
1K128K max
More room for conversation needs more memory.
One conversation, full GPU offload, FP16 KV cache.
YOUR MEMORY ESTIMATELive
7.2GiB
Recommended GPU memory
Room for your model. Space to think.
Model weights4.52 GiB
KV cache1.00 GiB
Runtime + headroom1.72 GiB
System RAM targetConservative loading allowance
16 GiBPlanning estimate, not a measured requirement.
02 Find your fit
Memory-compatible options Memory fits
UNIFIED MEMORY
Mac Studio · 64 GB
64 GB unified memory
Memory fits
ON-DEMAND COMPUTE
RunPod · A100 80 GB
80 GB VRAM
Provider links may be affiliate links. We may earn a commission at no extra cost to you. Memory fit does not guarantee runtime compatibility or performance.
What goes into the estimate?
Weight-file GB are converted to GiB (2³⁰ bytes). FP16 KV memory is 2 × layers × KV heads × head dimension × context tokens × 2 bytes. Runtime allowance is the larger of 1 GiB or 10% of weights. We then leave 10% of GPU capacity free.
How much system or Mac memory do I need?
System RAM is weights plus 8 GiB, rounded up to 8 GiB. Unified memory is the greater of GPU target plus 8 GiB or GPU target divided by 0.75, rounded up to 8 GiB. These are conservative loading and OS allowances, not measured peaks. Mac cards use this separate unified-memory target.
What should I check before buying?
Actual memory use depends on your runtime, model file, drivers, batching, and loading strategy. Catalog sizes use approximate parameters and effective bits per weight, except gpt-oss and Ministral 3 which use rounded published Ollama package sizes. Ministral 3 includes bundled vision weights, but image processing is not estimated and vLLM estimates are not yet verified. Gemma 3 and gpt-oss cache estimates allow full context per layer, without sliding-window savings. Compare the actual file and usable memory before buying. Hardware GB are conservatively treated as decimal bytes. CPU offload, image processing and multi-GPU performance are not modeled. vLLM options support FP16 and native gpt-oss MXFP4 and exclude Macs.
What about ChatGPT, Claude, Gemini and Codex?
Hosted AI services do the model processing remotely. Gemma and gpt-oss have downloadable weights for local use. Codex is a coding tool, and its memory needs depend on the model behind it. Compare local and hosted AI.