CanMyPCRunAILocal AI compatibility

What AI models can you run with 4GB of VRAM?

With 4GB of VRAM, 67 of 225 catalog builds have good/excellent reference fits at 4K context. 117 more run by offloading part of the model to system RAM, which works but is slower.

67

run well

117

with trade-offs

225

builds evaluated

Check my actual PC

Keep the first experiment small

With a 4 GB budget, start with a compact instruct build and a short context. Compare the 3B example below against its actual memory total; the download size alone can hide a cache or runtime allocation that exceeds the budget.

Quantization compresses weights; it does not remove context-cache or runtime costs. Compare the same model at Q4 and Q8 in the calculator, then validate its output on your task. Training and fine-tuning need a different memory budget.

Change system RAM and contextCompare exact builds8 GB vs 12 GB vs 16 GB

Four builds to compare by task

Editorial examples chosen for their documented tasks, all Q4_K_M at 4,096 tokens. Check the verdict before downloading. These are estimates, not benchmark rankings.

Short conversations and summaries

llama3.2:3b

A small instruction-tuned model for trying local chat with more memory headroom. Review summaries against the source text.

Task and build source · Catalog checked 2026-10-13

Good fit

Est. memory 2.9 GB · high confidence

Model fully fits in usable GPU VRAM.

ollama run llama3.2:3b

General instruction following

llama3.1:8b

An 8B instruction-tuned alternative to compare on your own prompts. It needs more memory than the 3B example and can still give incorrect answers.

Task and build source · Catalog checked 2026-10-13

Runs with trade-offs

Est. memory 5.8 GB · high confidence

Estimated 50% of layers offloaded to GPU (16/32); model does not fully fit in VRAM.

ollama run llama3.1:8b

Code explanation and editing

qwen2.5-coder:7b

A code-focused instruct build. Review generated changes and run your tests; a memory fit is not evidence that its code is correct.

Task and build source · Catalog checked 2026-10-13

Runs with trade-offs

Est. memory 5.3 GB · high confidence

Estimated 57% of layers offloaded to GPU (16/28); model does not fully fit in VRAM.

ollama run qwen2.5-coder:7b

A larger general-purpose comparison

qwen2.5:14b

A 14B instruct build for comparing answers when memory allows. Larger size does not establish better results for every task.

Task and build source · Catalog checked 2026-10-13

Runs with trade-offs

Est. memory 10.2 GB · high confidence

Estimated 25% of layers offloaded to GPU (12/48); model does not fully fit in VRAM.

ollama run qwen2.5:14b

After opening Ollama's prompt, enter /set parameter num_ctx 4096 to match this calculation, within the model's supported maximum. Install a current Ollama release that supports the build. Setup and troubleshooting.

Models that fit — largest memory first — for 4GB of VRAM

Heaviest first — each of these has a good/excellent reference fit with 4GB at 4K context, counting weights, KV cache, compute buffers and runtime overhead.

Models that fit in 4GB of VRAM
ModelBuildMemory neededVerdict
qwen3-embeddingqwen3-embedding:4bQ4_K_M~3.5 GBGood fit
dolphin-phidolphin-phi:2.7b-v2.6-q4_K_MQ4_K_MMemory scenario exceeds documented maximum of 2,048 tokens.~3.4 GBGood fit
phiphi:2.7b-chat-v2-q4_K_MQ4_K_MMemory scenario exceeds documented maximum of 2,048 tokens.~3.4 GBGood fit
granite3-guardiangranite3-guardian:2bQ8_0~3.4 GBGood fit
phi4-miniphi4-mini:3.8bQ4_K_M~3.4 GBGood fit
stablelm-zephyrstablelm-zephyr:3b-q4_K_MQ4_K_M~3.4 GBGood fit
stable-codestable-code:3b-code-q4_K_MQ4_K_M~3.4 GBGood fit
cogitocogito:3bQ4_K_M~3.1 GBGood fit
granite4.2granite4.2:3bQ4_K_M~3.0 GBGood fit
hermes3hermes3:3bQ4_K_M~2.9 GBGood fit
granite4.1granite4.1:3bQ4_K_M~2.8 GBGood fit
qwen3-vlqwen3-vl:2bQ4_K_M~2.7 GBExcellent fit
opencoderopencoder:1.5bQ4_K_M~2.7 GBExcellent fit
smallthinkersmallthinker:3b-preview-q4_K_MQ4_K_M~2.6 GBExcellent fit
shieldgemmashieldgemma:2bQ4_K_M~2.5 GBExcellent fit
gemma2gemma2:2b-text-q4_K_MQ4_K_M~2.5 GBExcellent fit
starcoder2starcoder2:3b-q4_K_MQ4_K_M~2.4 GBExcellent fit
exaone-deepexaone-deep:2.4bQ4_K_M~2.3 GBExcellent fit
exaone3.5exaone3.5:2.4bQ4_K_M~2.3 GBExcellent fit
granite3-densegranite3-dense:2bQ4_K_M~2.3 GBExcellent fit
granite3.1-densegranite3.1-dense:2bQ4_K_M~2.3 GBExcellent fit
granite3.3granite3.3:2bQ4_K_M~2.3 GBExcellent fit
granite3.2granite3.2:2bQ4_K_M~2.3 GBExcellent fit
granite3.2-visiongranite3.2-vision:2bQ4_K_M~2.3 GBExcellent fit
stablelm2stablelm2:1.6b-q4_K_MQ4_K_M~2.2 GBExcellent fit

How this was calculated

Every model is evaluated at 4K context against 4GB of VRAM, minus a reserve for the display and operating system, and assuming 32GB of system RAM with an illustrative 8-core/16-thread CPU and 500GB free disk. Available RAM and GPU budget/free memory are unknown. A working Vulkan backend on Windows is assumed; this is a capacity scenario, not a claim that all cards with this much memory have the same support. Your own machine may differ, which is what the PC scan measures directly.

4GB VRAM and local AI — frequently asked questions

Is 4GB of VRAM enough to run AI models locally?
Yes — 67 of the 225 builds evaluated have good/excellent reference fits with 4GB at 4K context, and 117 more run with part of the model offloaded to system RAM. The heaviest that fits is qwen3-embedding:4b, needing about 3.5 GB.
How should I choose a model for 4GB VRAM?
Choose for your task, then check memory. qwen3-embedding (qwen3-embedding:4b) is the heaviest catalog default that fits this reference budget; its size does not establish answer quality. Smaller models leave headroom for a longer context window, which often matters more than raw parameter count.
Which graphics cards have 4GB of VRAM?
Several cards ship with 4GB of VRAM across NVIDIA and AMD ranges.

Other VRAM amounts

6GB8GB10GB12GB16GB24GB32GB