Filter by your RAM
Enter RAM, CPU, and GPU to get an indicative estimate of memory and speed. The results are not benchmarks: check the model, the quantization, the context, and the free memory on your PC. The calculation runs in the browser.
Quick guide: which to choose
If you don't want to read everything, this is the part you need. Five rows, one per hardware tier.
| Your PC | RAM | Choose this | Why |
|---|---|---|---|
| PC with very limited memory | 2 GB | Example: 0.5-1B model | Runtime compatibility and context can make loading impractical. |
| PC with 4 GB | 4 GB | Example: 1-2B model | Leave memory for the OS, runtime, and context; results also depend on CPU instructions. |
| PC with 8 GB | 8 GB | Example: 3-8B model | A 7-8B Q4 can be at the limit; check free memory and runtime. |
| PC with 16 GB | 16 GB | Example: 7-14B model | Required memory changes with weights, quantization, and context. |
| PC with lots of RAM or a GPU | variable | Largest compatible model | Check memory, license, runtime, and quality on your task. |
| With a dedicated GPU | Variable VRAM | Model supported by the backend | VRAM limits which weights can be loaded; speed and compatibility depend on the card and the runtime. |
The rule of the two limits
Check whether the model fits in memory and whether the speed is adequate for your use. The table only offers examples: requirements and throughput depend on the file, runtime, CPU/GPU, context, and open applications.
Quantization explained
It's the concept that makes local AI possible on normal PCs. If you understand it, you also understand why 0.4 GB and 40 GB files exist for the same model.
A model stores its "weights" in a numeric format chosen at release time. Quantizing reduces the memory used, but it can change quality differently depending on the model, the format, and the task. Actual sizes also vary between files of the same model.
| Format | Bits per weight | Size of a 7B | Quality | When to use it |
|---|---|---|---|---|
| Q8_0 | 8 | varies | Reduces precision less, but isn't identical in every test | If available memory allows it |
| Q6_K | 6 | varies | Memory/precision trade-off | With headroom for context and system |
| Q5_K_M | 5 | varies | Intermediate trade-off; quality depends on the model | Compare on your desired task |
| Q4_K_M | 4 | varies | Common choice to reduce memory; the loss varies | Weigh memory and quality for your use |
| Q3_K_M | 3 | varies | More compact, with possibly bigger penalties | Try it if larger formats don't fit |
| Q2_K | 2 | varies | Most compressed format, potentially lower quality | Compare before choosing it |
Size and precision are different trade-offs
A larger model can achieve better results on some tasks, but it isn't a universal rule. Quantization can affect code, language, reasoning, and other tests differently. Compare specific versions on your use case instead of choosing only by parameters or bits.
Model size: what actually changes
A larger size generally needs more memory, but quality and speed depend on the model, the quantization, the task, and the hardware.
| Size | Indicative Q4 file | Indicative RAM | Speed | Use |
|---|---|---|---|---|
| 0.5-1B | about 0.5-1 GB | higher than the file | Depends on hardware | Short, simple tasks |
| 2-3B | about 1.5-2.5 GB | higher than the file | Depends on hardware | Light general use |
| 4B | about 2.5-4 GB | file + context | Depends on hardware | General use; check quality and language |
| 7-8B | about 4-6 GB | file + context | Depends on hardware | Can be at the limit on PCs with 8 GB total |
| 12-14B | about 8-11 GB | more than the file | Depends on hardware | Account for OS, context, and headroom |
| 24-32B | about 15-22 GB | more than the file | Depends on hardware | Adequate memory needed; some models are MoE |
| 70B | about 40 GB or more | much more than the file | Depends on hardware | Needs lots of memory; 24 GB VRAM is not enough for all weights |
Size drives memory and time
Generation repeatedly works over the weights and the context data, so larger models tend to need more memory and time. Real throughput depends on bandwidth, CPU/GPU, backend, quantization, and context length: it can't be inferred precisely from size alone.
Models for specific tasks
A specialized model can be better suited to a task, but quality depends on the version and the request. Try the model on the work you want to do.
Qwen2.5-Coder 7B
A coding-oriented family. Available variants and quantizations change; compare the model card with your task and runtime.
DeepSeek-R1-Distill-Qwen-7B
A distilled variant of the DeepSeek-R1 family. Quality varies by task and version; check the model's requirements and behavior before using it.
Qwen2.5 7B Instruct
Instruct model of the Qwen family. It can be used for text and writing; compare Italian quality with prompts representative of your case.
Gemma 3 4B
Multimodal variant that can process images with a compatible runtime. Check size, memory, backend, and how the application handles your data.
Qwen3 30B-A3B
MoE model with about 3 billion active parameters per token. Loading it still requires memory for the full weights, the runtime, and the context; speed and quality depend on hardware, quantization, and backend. It isn't a universal choice: verify the requirements of the exact variant.
Granite 3.3 2B Instruct
Instruct variant of the Granite family. Features, formats, and terms of use must be checked in the model card and in the license of the specific artifact.
Full table
All models in the catalog with their technical data. The CPU threshold is an internal heuristic index, not a manufacturer requirement nor a guaranteed speed measurement.
| Model | Task | Parameters | Q4 file | RAM | Estimated CPU threshold | Why choose it |
|---|
Licenses and commercial use
The license can vary across families, versions, and files distributed by third parties. For commercial use, check the terms attached to the specific model and the specific quantization, including any obligations and restrictions.
Check the source of the file you download
The model card must state author, license, terms of use, and origin. Don't infer the license from the family name alone or from the app used to download it. If the use matters to a business, keep a copy of the applicable terms and ask for qualified advice when needed.
Frequently asked questions
What is the best model for 8 GB of RAM?
With 8 GB total, free memory depends on the OS, applications, and context. Small or mid-sized quantized models are a starting point; some 7-8B models can be at the limit. The examples in the catalog are not an up-to-date ranking: verify the file, runtime, and requirements before downloading.
What's the difference between a 7B and a 14B?
The result varies with the model and the task: more parameters don't guarantee a better answer. A quantized 14B generally needs more memory than a 7B; also consider context, runtime, and operating system. Check the actual size and leave additional room.
What is Q4_K_M quantization and does it really matter?
Quantization reduces the memory used by the weights compared with higher precisions. Q4_K_M is a common trade-off, but the impact on quality varies by model and task; compare the specific variant. More compressed formats require less memory but can degrade some capabilities more.
Does a larger model always give better results?
No. Size, quality, and speed don't have a simple relationship: the result depends on the model, task, quantization, and hardware. More parameters often increase memory and time, but quality must be judged on your use case.
Which model speaks Italian best?
There is no universal winner: language, version, prompt, and task change the results. Qwen, Llama, Gemma, and other families have instruct variants that can reply in Italian; try the available versions with your own requests.
How do I free memory for a larger model?
Close unneeded applications, shorten the context, and choose a lighter quantization when available. A lightweight operating system can help, but it doesn't guarantee that a model will fit: consider cache, buffers, and real free memory.
Where do I download them?
From Hugging Face, which is the reference repository, or directly from LM Studio or with an Ollama command. The LM Studio configurator shows which format to download for your hardware; if you prefer the terminal and want to serve models to other programs, the guide is on the Ollama vs LM Studio page.