RunsOnMy ModelsGPUs & MacsBy VRAMBest GPUCan I run…?API pricesVRAM calculatorNew

Can I run gpt-oss-120b on my PC?

gpt-oss 120B (117B, 5.8B active) needs about 59.5 GB of GPU memory at 4-bit with an 8K context. No single consumer GPU we track runs it fully in VRAM at 4-bit. With llama.cpp offload, the smallest card that can run it is the RX 7600 with 64 GB of system RAM, at an estimated 12.0–22.9 tokens/sec (usable for chat). The smallest Mac that runs it is the Mac M2 Max 96GB.

59.5 GBneeded at 4-bit, 8K
8GPUs/Macs: fully in memory
33GPUs: with CPU offload
M2 Max 96GBsmallest Mac

gpt-oss 120B on every GPU and Mac

Best quant that fits fully in memory at 8K context with a speed ceiling; otherwise the llama.cpp offload plan (system RAM needed, estimated speed from DDR4-3200 to DDR5-6000). Macs and unified-memory PCs share one pool, so offload adds nothing there. How offload is estimated.

HardwareMemoryVerdictDetails
RX 76008 GBOffload Q4_K_M, 64 GB RAM · ~12–23 tok/sdetails
RTX 40608 GBOffload Q4_K_M, 64 GB RAM · ~12–23 tok/sdetails
RTX 50608 GBOffload Q4_K_M, 64 GB RAM · ~12–24 tok/sdetails
RTX 3060 Ti8 GBOffload Q4_K_M, 64 GB RAM · ~12–24 tok/sdetails
RTX 4060 Ti 8GB8 GBOffload Q4_K_M, 64 GB RAM · ~12–23 tok/sdetails
RTX 30708 GBOffload Q4_K_M, 64 GB RAM · ~12–24 tok/sdetails
RTX 3080 10GB10 GBOffload Q4_K_M, 64 GB RAM · ~13–26 tok/sdetails
Arc B58012 GBOffload Q4_K_M, 64 GB RAM · ~13–25 tok/sdetails
RTX 3060 12GB12 GBOffload Q4_K_M, 64 GB RAM · ~13–25 tok/sdetails
RX 7700 XT12 GBOffload Q4_K_M, 64 GB RAM · ~13–25 tok/sdetails
RTX 507012 GBOffload Q4_K_M, 64 GB RAM · ~13–27 tok/sdetails
RTX 407012 GBOffload Q4_K_M, 64 GB RAM · ~13–26 tok/sdetails
RTX 4070 Super12 GBOffload Q4_K_M, 64 GB RAM · ~13–26 tok/sdetails
RTX 4070 Ti12 GBOffload Q4_K_M, 64 GB RAM · ~13–26 tok/sdetails
RTX 3080 12GB12 GBOffload Q4_K_M, 64 GB RAM · ~13–27 tok/sdetails
Arc A770 16GB16 GBOffload Q4_K_M, 64 GB RAM · ~14–27 tok/sdetails
RTX 5060 Ti 16GB16 GBOffload Q4_K_M, 64 GB RAM · ~14–27 tok/sdetails
RTX 4060 Ti 16GB16 GBOffload Q4_K_M, 64 GB RAM · ~13–25 tok/sdetails
RX 7800 XT16 GBOffload Q4_K_M, 64 GB RAM · ~14–28 tok/sdetails
RX 7900 GRE16 GBOffload Q4_K_M, 64 GB RAM · ~14–28 tok/sdetails
RX 907016 GBOffload Q4_K_M, 64 GB RAM · ~14–28 tok/sdetails
RX 9070 XT16 GBOffload Q4_K_M, 64 GB RAM · ~14–28 tok/sdetails
RTX 5070 Ti16 GBOffload Q4_K_M, 64 GB RAM · ~14–29 tok/sdetails
RTX 4070 Ti Super16 GBOffload Q4_K_M, 64 GB RAM · ~14–28 tok/sdetails
RTX 4080 Super16 GBOffload Q4_K_M, 64 GB RAM · ~14–28 tok/sdetails
RTX 508016 GBOffload Q4_K_M, 64 GB RAM · ~14–29 tok/sdetails
RTX 408016 GBOffload Q4_K_M, 64 GB RAM · ~14–28 tok/sdetails
RX 7900 XT20 GBOffload Q4_K_M, 64 GB RAM · ~16–31 tok/sdetails
RX 7900 XTX24 GBOffload Q4_K_M, 64 GB RAM · ~17–34 tok/sdetails
RTX 309024 GBOffload Q4_K_M, 64 GB RAM · ~17–34 tok/sdetails
RTX 409024 GBOffload Q4_K_M, 64 GB RAM · ~17–34 tok/sdetails
RTX 3090 Ti24 GBOffload Q4_K_M, 64 GB RAM · ~17–34 tok/sdetails
RTX 509032 GBOffload Q4_K_M, 64 GB RAM · ~22–44 tok/sdetails
DGX Spark128 GBF16 · ~85 tok/sdetails
Ryzen AI Max+ 395 128GB128 GBF16 · ~80 tok/sdetails
Mac M1 16GB16 GBNodetails
Mac M4 16GB16 GBNodetails
Mac M2 24GB24 GBNodetails
Mac M3 24GB24 GBNodetails
Mac M4 Pro 24GB24 GBNodetails
Mac M4 32GB32 GBNodetails
Mac M1 Pro 32GB32 GBNodetails
Mac M3 Pro 36GB36 GBNodetails
Mac M4 Pro 48GB48 GBNodetails
Mac M4 Pro 64GB64 GBNodetails
Mac M1 Max 64GB64 GBNodetails
Mac M4 Max 64GB64 GBNodetails
Mac M2 Max 96GB96 GBF16 · ~124 tok/sdetails
Mac M3 Max 128GB128 GBF16 · ~124 tok/sdetails
Mac M4 Max 128GB128 GBF16 · ~170 tok/sdetails
Mac M2 Ultra 192GB192 GBF16 · ~249 tok/sdetails
Mac M3 Ultra 256GB256 GBF16 · ~255 tok/sdetails
Mac M3 Ultra 512GB512 GBF16 · ~255 tok/sdetails

Questions

How much VRAM do I need for gpt-oss 120B?

About 59.5 GB at 4-bit (Q4_K_M) with an 8K context, including KV cache and runtime overhead. Longer contexts need more; the model page has the full table by quant and context.

Why can a mixture-of-experts model run on a smaller card?

Only 5.8B of its 117B parameters are used per token. llama.cpp's --n-cpu-moe keeps the expert weights of some layers in system RAM while attention and the KV cache stay on the GPU, so each token reads far less from slow RAM than a dense model of the same size would.

Other popular checks: Can I run gpt-oss-20b?Can I run Llama 70B?Can I run Qwen3?Can I run DeepSeek R1? · Best GPU by budget · Mac vs PC

Model data: Hugging Face Hub API and GGUF file sizes. GPU specs: vendor pages. Speeds are estimates, not benchmarks. Methodology · Report an error.