Can your gaming laptop run local AI? The best LLMs for laptop GPUs

Most gaming laptops sold in the last few years can run AI models locally. The question is which ones, and how fast. Laptop GPUs share names with desktop cards but usually have less memory and less bandwidth, so a "4070" in a laptop is a very different tool from a 4070 in a desktop. This guide shows what each common laptop GPU can run.

All numbers come from our checker at 8K context, assuming 32 GB of system RAM for models that spill over.

Laptop GPUs are not the same as desktop ones

• Less VRAM. The RTX 4070 Laptop has 8 GB; the desktop 4070 has 12 GB. The RTX 4090 Laptop has 16 GB; the desktop 4090 has 24 GB.
• Less bandwidth. The RTX 4060 Laptop reads memory at 256 GB/s, close to the desktop 4060 (272 GB/s), but the 4090 Laptop manages 576 GB/s against 1,008 GB/s on the desktop card. Bandwidth decides speed.
• Power limits. Laptops run their GPUs at lower wattage, and two laptops with the same GPU can differ a lot. Treat the speeds below as a realistic ceiling, and plug in the charger when you run models.

What each laptop GPU can run

Two surprises stand out. The older RTX 3060 Laptop is faster than the newer RTX 4050 Laptop because it has more memory bandwidth (336 vs 192 GB/s). And the RTX 4060 and 4070 Laptop run local models at the same speed, because both have 8 GB on the same 128-bit memory bus. If you are buying a laptop for local AI, the 4070 Laptop is not worth the extra money over the 4060.

6 GB laptops: small but useful

With about 5.4 GB usable, 6 GB laptops run the strong 4B models: Qwen3.5 4B, Qwen3 4B and Gemma 3 4B. They handle summaries, emails, simple coding questions and quick chat well. 7B and 8B models do not fit, even at low quantization. Our guide to 4 GB and 6 GB GPUs goes deeper.

8 GB laptops: the 8B sweet spot

The RTX 4060 and 4070 Laptop fit every popular 7B to 9B model at Q4_K_M: Qwen3 8B for general use, Qwen3.5 9B if you want image understanding, Qwen2.5-Coder 7B for code and DeepSeek-R1 Distill 8B for step-by-step reasoning. Around 30 tokens per second feels fast. See the best models for 8 GB for picks by task.

12 and 16 GB laptops

The RTX 4080 Laptop moves you up to 12B models such as Gemma 4 12B and Mistral Nemo 12B, or 8B models at full Q8 quality. The RTX 4090 Laptop is the only laptop GPU that fits gpt-oss 20B and the 14B models at Q5, which puts it in the same class as a desktop 16 GB card. Read more in 12 GB and 16 GB guides.

The trick for small laptops: mixture-of-experts models

Models like Qwen3 30B-A3B and gpt-oss 20B are too big for any laptop GPU below 16 GB, but they only use a few billion parameters for each word. With 32 GB of system RAM they still run at roughly 9 to 12 tokens per second on every laptop GPU in the table, with part of the model in RAM. That is slower than a model that fits, but it gives a 6 or 8 GB laptop access to much smarter answers. How mixture-of-experts models work explains why.

What about MacBooks?

MacBooks with Apple Silicon work differently: the GPU shares the whole unified memory, so a MacBook Pro with 36 or 48 GB runs models no Windows laptop GPU can hold. See the best models for every Mac.

Tips for running models on a laptop

1. Plug in. On battery, most laptops cut GPU power heavily.
2. Use the dedicated GPU. In the NVIDIA Control Panel or Windows graphics settings, set Ollama or LM Studio to use the high-performance GPU.
3. Watch the heat. Long sessions push laptops to their thermal limits; a cooling pad helps keep speeds steady.
4. Keep context short. Every extra 1,000 tokens of conversation uses VRAM you do not have much of. See how context length uses memory.

Find your exact laptop GPU in the GPU list, or pick it in the checker to see every model it can run.