LocalLLMChecker

Direct answers

Can my GPU run this local AI model?

One page per pairing worth asking about, each with a yes or a no at the top, the memory and speed underneath, and the settings we would use. When the answer is no, the page says what to run instead and what it would take to fix.

We only publish a page when the pairing is a real question. A 70B model on a 6 GB card is not, so it has no page, and the device's own page lists everything it can and cannot run. A page can say yes, partly or no. Partly means that some of the model sits in system RAM, or that a setting such as the macOS memory limit has to change. Each answer comes from the same engine as the calculator, which you can use for any model and any device we track, and the engine is checked against a real machine on the validation page.

GeForce RTX 3050 6GB

6 GB, 168 GB/s, Desktop GPU. 7 of 10 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct does not fit Qwen3 14B does not fit Phi 4 partly GPT OSS 20B runs Qwen3 30B A3B runs GPT OSS 120B runs

GeForce RTX 3060 12GB

12 GB, 360 GB/s, Desktop GPU. 10 of 13 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B does not fit Gemma 4 31B Instruct partly Qwen3 32B does not fit Qwen3 30B A3B runs GPT OSS 120B runs

GeForce RTX 4060

8 GB, 272 GB/s, Desktop GPU. 8 of 11 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B does not fit Phi 4 does not fit GPT OSS 20B runs Qwen3.5 27B partly Qwen3 30B A3B runs GPT OSS 120B runs

GeForce RTX 4060 Ti 16GB

16 GB, 288 GB/s, Desktop GPU. 11 of 13 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct does not fit Qwen3 32B does not fit Qwen3 30B A3B runs GPT OSS 120B runs

GeForce RTX 4070

12 GB, 504 GB/s, Desktop GPU. 10 of 13 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B does not fit Gemma 4 31B Instruct partly Qwen3 32B does not fit Qwen3 30B A3B runs GPT OSS 120B runs

GeForce RTX 4070 Ti SUPER

16 GB, 672 GB/s, Desktop GPU. 11 of 13 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct does not fit Qwen3 32B does not fit Qwen3 30B A3B runs GPT OSS 120B runs

GeForce RTX 5070 Ti

16 GB, 896 GB/s, Desktop GPU. 11 of 13 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct does not fit Qwen3 32B does not fit Qwen3 30B A3B runs GPT OSS 120B runs

GeForce RTX 5080

16 GB, 960 GB/s, Desktop GPU. 11 of 13 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct does not fit Qwen3 32B does not fit Qwen3 30B A3B runs GPT OSS 120B runs

GeForce RTX 3090

24 GB, 936 GB/s, Desktop GPU. 13 of 14 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct runs Qwen3 32B runs Qwen3 30B A3B runs Llama 3.3 70B Instruct does not fit GPT OSS 120B runs

GeForce RTX 4090

24 GB, 1008 GB/s, Desktop GPU. 13 of 14 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct runs Qwen3 32B runs Qwen3 30B A3B runs Llama 3.3 70B Instruct does not fit GPT OSS 120B runs

GeForce RTX 5090

32 GB, 1792 GB/s, Desktop GPU. 13 of 14 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct runs Qwen3 32B runs Qwen3 30B A3B runs Llama 3.3 70B Instruct does not fit GPT OSS 120B runs

Radeon RX 7900 XTX

24 GB, 960 GB/s, Desktop GPU. 13 of 14 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct runs Qwen3 32B runs Qwen3 30B A3B runs Llama 3.3 70B Instruct does not fit GPT OSS 120B runs

Radeon RX 9070 XT

16 GB, 645 GB/s, Desktop GPU. 11 of 13 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct does not fit Qwen3 32B does not fit Qwen3 30B A3B runs GPT OSS 120B runs

Arc B580

12 GB, 456 GB/s, Desktop GPU. 10 of 13 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B does not fit Gemma 4 31B Instruct partly Qwen3 32B does not fit Qwen3 30B A3B runs GPT OSS 120B runs

Apple M4 16GB

16 GB, 120 GB/s, Unified memory. 9 of 12 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct partly Qwen3 32B partly Qwen3 30B A3B partly

Apple M4 Pro 48GB

48 GB, 273 GB/s, Unified memory. 13 of 14 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct runs Qwen3 32B runs Qwen3 30B A3B runs Llama 3.3 70B Instruct runs GPT OSS 120B does not fit

Apple M4 Max 64GB

64 GB, 546 GB/s, Unified memory. 13 of 14 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct runs Qwen3 32B runs Qwen3 30B A3B runs Llama 3.3 70B Instruct runs GPT OSS 120B does not fit

Apple M4 Max 128GB

128 GB, 546 GB/s, Unified memory. 14 of 14 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct runs Qwen3 32B runs Qwen3 30B A3B runs Llama 3.3 70B Instruct runs GPT OSS 120B runs

Apple M3 Max 36GB

36 GB, 400 GB/s, Unified memory. 12 of 14 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct runs Qwen3 32B runs Qwen3 30B A3B runs Llama 3.3 70B Instruct partly GPT OSS 120B does not fit

Ryzen AI Max+ 395 128GB

128 GB, 256 GB/s, Unified memory. 14 of 14 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct runs Qwen3 32B runs Qwen3 30B A3B runs Llama 3.3 70B Instruct runs GPT OSS 120B runs

GeForce RTX 3070

8 GB, 448 GB/s, Desktop GPU. 8 of 11 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B does not fit Phi 4 does not fit GPT OSS 20B runs Qwen3.5 27B partly Qwen3 30B A3B runs GPT OSS 120B runs

GeForce RTX 4090 Laptop

16 GB, 576 GB/s, Laptop GPU. 11 of 13 pairings run at a quantization worth using.

Qwen2.5 7B Instruct runs Llama 3.1 8B Instruct runs Qwen3 8B runs Qwen3.5 9B runs Gemma 4 12B Instruct runs Qwen3 14B runs Phi 4 runs GPT OSS 20B runs Qwen3.5 27B runs Gemma 4 31B Instruct does not fit Qwen3 32B does not fit Qwen3 30B A3B runs GPT OSS 120B runs