What Local LLM Can My Computer Run?

Before you begin, there are three key steps:

1. Check your hardware specs (operating system, memory, and graphics card details).
2. Pick a recommended model based on what your computer can handle.
3. Use the calculator link in each example to see if the selected model fits your hardware before downloading.

Start with a small model that leaves room for your other work. An 8GB computer can try a tiny text model; a 16GB computer has more useful choices; a PC with a dedicated graphics card must also meet that card's separate memory limit. The eight examples below give you an exact model file, app, and calculator link, so you can change the settings to match your machine before downloading.

If you do not know your hardware yet, start with Hardware Fit Lab. If you know your RAM and graphics memory, open the RAM and VRAM calculator. Both use the answers you enter; neither scans your computer or runs a model.

Find the three facts that decide your starting point

Write down your operating system, installed memory, and graphics processor. On Windows, check Settings → System → About for RAM and processor, then Task Manager → Performance for the GPU and its dedicated GPU memory. On a Mac, use Apple menu → About This Mac to find the chip and memory. The Mac examples here mean Apple Silicon, not an Intel Mac.

A PC with 16GB RAM and 6GB VRAM does not have 22GB of interchangeable model memory. Likewise, Windows' “shared GPU memory” isn't extra dedicated VRAM. If you cannot establish the dedicated capacity, leave it unknown rather than choosing “no dedicated GPU.”

The examples use the familiar advertised labels 8GB, 16GB and 32GB for computers. The calculator interprets entered memory capacities as GiB and reports GiB. Model download sizes below use decimal GB; for example, the 2.497GB Qwen3 4B download occupies about 2.33GiB before conversation and app memory are added.

Memory is only the first check. LM Studio requires AVX2 on Windows x64, recommends at least 16GB RAM, and requires Apple Silicon with macOS 14 or newer on Mac. Ollama requires Windows 10 22H2 or newer, or macOS 14 or newer. A GPU with enough VRAM still needs support from the chosen app and its drivers. The GPU examples below assume that support is established; capacity alone cannot establish it.

Choose an exact file, not just a model name

“4B” means roughly four billion model parameters. It is not a download size. GGUF is a model file format; Q4_K_M and Q8_0 describe how its weights are compressed, also called quantization. Different compressions of the same model need different amounts of memory and can give different answers.

These examples use four text models from Qwen's own GGUF repositories:

These are a practical shortlist, not a performance ranking. A bigger model deserves the extra memory only if it improves your own task. For writing, compare two rewrites of the same paragraph. For basic coding, ask each model to explain and test one small function. For chat, ask a question whose answer you can check. The tiny 0.6B model is a learning exercise, not a substitute for checking consequential facts.

The linked Qwen repositories publish these files under Apache 2.0. Read the included license when distributing models or incorporating them into a product. The filenames matter: a different uploader's Q4 file, an MLX download, or an Ollama library tag may represent a different file size. Do not substitute one while keeping the same estimate.

The settings behind all eight examples

Every baseline below uses 4,096 tokens of context, one user, text chat, the app named in the example, and the calculator's existing 20% safety allowance. Conversation memory uses the calculator's Auto setting, which allows for FP16; Q4 model weights do not mean the conversation memory is also Q4. CPU offload is off in every baseline.

Context is the space for instructions, the conversation, any supplied passages, and the answer together. A 4,096-token setting does not leave 4,096 tokens exclusively for a document. Keep a first prompt short and leave room for the response. These settings are for learning and focused tasks, not large coding agents or whole case files.

The calculator first leaves memory for the operating system and, on a dedicated GPU, the display. It then estimates model weights, conversation memory, app memory, temporary calculations, and the safety allowance. “Remaining” below means after those allowances. It does not mean your browser and every other running app have been measured. The model-weight estimates use the specified download sizes. Conversation-memory estimates use the recorded architecture dimensions; app and temporary-memory use also vary with the runner and settings. Treat a narrow margin as a reason to start smaller. The estimates cannot tell you how many words per second your particular computer will produce.

The worked figures below use two decimal places. The calculator's result cards round to one decimal place, so a result card can show 8.9GiB where the example says 8.87GiB. Both round the same underlying estimate; the extra digit is for comparison, not a claim of measured precision.

Eight computers, with settings you can reopen

1. Windows with 8GB RAM and no dedicated GPU

Try Qwen3 0.6B, Q8_0, in Ollama. Enter 8GiB RAM and no dedicated GPU. The estimate leaves 4.00GiB available after the operating-system allowance and needs about 3.14GiB RAM, leaving 0.86GiB. That is a tight start, suitable for a short first conversation with other heavy apps closed.

Open the 8GB Windows calculation.

Alternatively, postpone larger models until you have more RAM; this keeps the computer usable but limits what the model can handle. Reducing context can help memory, but it also shortens what you can ask. Follow Install Ollama, then use the exact model command below instead of a larger starter model from another guide. CPU speed remains a separate constraint.

2. Windows with 16GB RAM and integrated graphics

Start with Qwen3 4B, Q4_K_M, in LM Studio. Treat the integrated graphics as sharing system RAM; do not invent dedicated VRAM. The calculator needs about 5.97GiB RAM out of 12.00GiB available, leaving 6.03GiB.

Open the 16GB integrated-graphics calculation.

This is a memory fit, not a promise of GPU acceleration. Plan for CPU use unless the app supports your specific integrated GPU. If answers arrive too slowly, try the 0.6B file to separate speed limits from installation problems; its smaller size comes with a substantial capability tradeoff. Use Install LM Studio and check the Windows processor requirement before downloading.

3. Windows with 16GB RAM and a 6GB dedicated GPU

Start with Qwen3 4B, Q4_K_M, in LM Studio. The estimate needs 3.81GiB VRAM out of 5.25GiB available, leaving 1.44GiB VRAM. Separately, it allows 4.00GiB system RAM out of 12.00GiB, leaving 8.00GiB RAM.

Open the 16GB RAM / 6GB VRAM calculation.

The useful alternative is a smaller model if you need more context or other GPU applications open. Choosing an 8B model instead increases GPU demand; spare system RAM doesn't automatically solve that. Follow the Windows steps in Install LM Studio, then confirm the app uses your supported dedicated GPU.

4. Windows with 32GB RAM and an 8GB dedicated GPU

Qwen3 8B, Q4_K_M, in LM Studio is a tight GPU fit. It needs about 6.71GiB VRAM out of 7.25GiB available, leaving only 0.54GiB VRAM. The separate system-RAM estimate is 4.00GiB out of 24.96GiB, leaving 20.96GiB RAM.

Open the 32GB RAM / 8GB VRAM calculation.

Start with Qwen3 4B instead if you want more GPU room before experimenting. The tradeoff is a smaller model, while the benefit is less pressure from conversation memory and other graphics work. Follow the Install LM Studio instructions. Keep the 4,096-token setting for the first trial: the context example below shows why a longer conversation changes this answer.

5. Windows with 32GB RAM and a 12GB dedicated GPU

Check Qwen3 14B, Q4_K_M, in LM Studio before assuming it fits. At these settings, it needs about 11.36GiB VRAM, but the calculator leaves 11.04GiB available: a 0.32GiB VRAM shortfall. The separate 4.00GiB RAM requirement fits within 24.96GiB. That spare RAM does not turn this into a GPU-only fit.

Open the 32GB RAM / 12GB VRAM calculation.

Choose Qwen3 8B Q4_K_M for the first setup instead. Open that smaller-model calculation. It needs about 6.71GiB VRAM, leaving 4.33GiB in this GPU estimate. That sacrifices model size for useful working room. Partial CPU offload is another option to investigate in the app, but it changes the execution path and can slow answers; it is not assumed here. Use Install LM Studio.

6. Apple Silicon Mac with 8GB unified memory

Try Qwen3 0.6B, Q8_0, in Ollama. The model and app need about 3.14 GiB of unified memory out of 4.00 GiB available after the operating-system allowance, leaving 0.86 GiB. All the model's GPU and CPU work competes with your other Mac apps for this same pool.

Open the 8GB Mac calculation.

LM Studio is an alternative if you prefer its model browser, but its larger app-memory allowance leaves less room. Do not jump to a 4B model just because its download is smaller than 8GB. Follow Install Ollama on Mac; check macOS 14 or newer and keep the first trial to short text.

7. Apple Silicon Mac with 16GB unified memory

Try Qwen3 8B, Q4_K_M, in LM Studio. The estimate needs 8.87GiB unified memory out of 12.00GiB available, leaving 3.13GiB. This gives more room for everyday writing or short coding conversations than the 8GB example.

Open the 16GB Mac calculation.

Choose Qwen3 4B Q4_K_M instead if you need to keep a busy browser and other large apps open. It uses about 5.97GiB in the same calculation, leaving 6.03GiB. Compare whether the 8B model improves your actual prompt enough to justify the extra memory. Follow Install LM Studio on Mac.

The 16GB Mac example leaves about 3.1GiB unified memory after the stated allowances. The remaining capacity still depends on what else you run.

8. Apple Silicon Mac with 32GB unified memory

Qwen3 14B, Q4_K_M, in LM Studio is a reasonable next comparison. It needs about 14.13 GiB of unified memory out of 26.24 GiB available, leaving 12.11 GiB. This is why the same 14B file can fit this Mac estimate while missing the separate 12GB GPU limit in example 5.

Open the 32GB Mac calculation.

The 8B file is still a useful alternative: choose it if the 14B model adds waiting without a meaningful improvement in your writing or coding task. Capacity doesn't determine response speed, and these figures don't include a second loaded model or a document-search service. Follow Install LM Studio on Mac.

What changes when you increase context?

Take example 4: the PC with 32GB RAM and an 8GB GPU. Keep the same Qwen3 8B Q4_K_M file, LM Studio, one user, Auto conversation-memory precision, and 20% allowance. Change only the context from 4,096 to 32,768 tokens.

Open the longer-context calculation.

The model download has not grown. Qwen3 8B has 36 layers, 8 KV heads and head dimension 128. At two bytes per FP16 value, the K+V cache is 2 × 36 × 8 × 128 × tokens × 2 / 1024³ GiB for one user. The conversation-memory estimate rises from 0.5625GiB to 4.5GiB, and the temporary-memory allowance also grows. A model's published context limit says what it can accept with suitable settings and hardware; it does not mean your GPU can afford that context. Return to 4,096 tokens or choose a smaller model before considering offload. Do not remove the safety allowance to make the result look better.

Increasing only context changes the GPU result to insufficient memory. The separate system RAM surplus does not remove that VRAM shortfall.

An unknown GPU is a different answer from insufficient memory

Suppose you know the PC has 32GB RAM, but you do not know its dedicated VRAM. For Qwen3 8B Q4_K_M in LM Studio at 4,096 tokens, the calculator can still estimate 6.71GiB VRAM needed. It cannot compare that to a capacity you haven't supplied.

Open the calculation with VRAM not supplied.

Read “Estimate only” as a request for the missing fact. Find the dedicated capacity, enter it, and then read the new result. Do not read it as a fit, and do not replace the unknown with zero unless you know you are planning a CPU-only run.

The GPU requirement can be estimated while the GPU capacity remains unknown. Enter the missing capacity before treating this as a fit decision.

Turn the estimate into a first conversation

1. Install the chosen app using the linked guide. Check its OS and processor requirements. Keep disk space available for the app and the model download; disk storage is separate from RAM.
2. Select the exact file. In LM Studio, search the named Qwen/Qwen3-4B-GGUF, Qwen/Qwen3-8B-GGUF, or Qwen/Qwen3-14B-GGUF repository and choose Q4_K_M. If importing a downloaded GGUF instead, confirm the filename against the table.
3. Set context to 4,096 before the first conversation. In Ollama, use its context setting; in LM Studio, use the model's load settings. Changing this website's calculator does not change either app.
4. Start one short text conversation. For example: “Rewrite this paragraph in two shorter sentences without adding facts,” followed by a paragraph you wrote. For Qwen3, you can append /no_think for a simple first trial. Avoid long reasoning requests that need a much larger answer allowance.
5. Check the actual computer. Look for a normal answer, a usable desktop, and an acceptable wait time. On Windows, watch RAM and GPU usage in Task Manager; on Mac, watch Memory Pressure and swap in Activity Monitor. If the computer becomes unresponsive, unload the model, close heavy apps, and try a smaller model or context. If it still fails, preserve the exact error and the log entry at that time. Follow memory failures for allocation errors, GPU detection when acceleration is absent, or connection checks when the interface cannot reach the runner. Update only when the error identifies a version or compatibility requirement.

For the 0.6B Ollama examples, this command names the exact official GGUF and compression:

ollama run hf.co/Qwen/Qwen3-0.6B-GGUF:Q8_0

In the interactive Ollama conversation, set the example context explicitly:

/set parameter num_ctx 4096

After a response, ollama ps shows the loaded model, processor allocation, and context. That is a useful check of what the runner actually loaded; the calculator link alone cannot establish it.

Open a worked-example link, change it to your hardware, and copy its address from the browser if you want to reopen those inputs later. Reopening restores calculation settings, not an installed model, saved conversation, or permanent historical result. If you change the app, model file, context, users, or offload, recalculate before relying on the earlier answer.

When to stop increasing model size

Stop when a larger model adds memory pressure or wait time without improving your real task. Larger models can be significantly slower to respond, especially on computers with less powerful processors or limited memory. Run times vary by hardware, and the fastest model on one computer might not be the fastest on another. A model that loads isn't necessarily useful. If a tiny prompt succeeds but document work fails, document extraction, search, retrieved text, and longer context may create a different workload. Use Chat with PDFs locally before treating a chat-memory estimate as a plan for confidential files.

If a model won't load, a GPU isn't used, or answers arrive too slowly, follow Troubleshooting. For the detailed memory components, see Local LLM RAM and VRAM requirements. Keep your exact model file and settings beside the problem description: “Qwen3 8B Q4_K_M at 32,768 tokens” is much more useful than “an 8B model is broken.”

Sources

• Qwen3 8B architecture configuration.
• Qwen's official GGUF files and model instructions: Qwen3 0.6B, Qwen3 4B, Qwen3 8B, and Qwen3 14B.
• LM Studio system requirements and model download instructions.
• Ollama's Windows requirements, Mac requirements, context and processor-allocation instructions, and interactive context command.