Can I run gpt-oss-20b on my PC?
gpt-oss 20B (21B, 4.2B active) needs about 11.8 GB of GPU memory at 4-bit with an 8K context. The cheapest consumer GPU in our table that runs it fully in VRAM at 4-bit or better is the Arc A770 16GB (16 GB, $349 launch MSRP). With llama.cpp offload, the smallest card that can run it is the RX 7600 with 32 GB of system RAM, at an estimated 27.0–45.3 tokens/sec (usable for chat). The smallest Mac that runs it is the Mac M4 Pro 24GB.
gpt-oss 20B 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.
| Hardware | Memory | Verdict | Details |
|---|---|---|---|
| RX 7600 | 8 GB | Offload Q4_K_M, 32 GB RAM · ~27–45 tok/s | details |
| RTX 4060 | 8 GB | Offload Q4_K_M, 32 GB RAM · ~27–44 tok/s | details |
| RTX 5060 | 8 GB | Offload Q4_K_M, 32 GB RAM · ~29–52 tok/s | details |
| RTX 3060 Ti | 8 GB | Offload Q4_K_M, 32 GB RAM · ~29–52 tok/s | details |
| RTX 4060 Ti 8GB | 8 GB | Offload Q4_K_M, 32 GB RAM · ~27–45 tok/s | details |
| RTX 3070 | 8 GB | Offload Q4_K_M, 32 GB RAM · ~29–52 tok/s | details |
| RTX 3080 10GB | 10 GB | Offload Q4_K_M, 32 GB RAM · ~49–85 tok/s | details |
| Arc B580 | 12 GB | Offload Q4_K_M, 32 GB RAM · ~72–98 tok/s | details |
| RTX 3060 12GB | 12 GB | Offload Q4_K_M, 32 GB RAM · ~63–83 tok/s | details |
| RX 7700 XT | 12 GB | Offload Q4_K_M, 32 GB RAM · ~70–94 tok/s | details |
| RTX 5070 | 12 GB | Offload Q4_K_M, 32 GB RAM · ~85–124 tok/s | details |
| RTX 4070 | 12 GB | Offload Q4_K_M, 32 GB RAM · ~75–104 tok/s | details |
| RTX 4070 Super | 12 GB | Offload Q4_K_M, 32 GB RAM · ~75–104 tok/s | details |
| RTX 4070 Ti | 12 GB | Offload Q4_K_M, 32 GB RAM · ~75–104 tok/s | details |
| RTX 3080 12GB | 12 GB | Offload Q4_K_M, 32 GB RAM · ~94–146 tok/s | details |
| Arc A770 16GB | 16 GB | Q8_0 · ~231 tok/s | details |
| RTX 5060 Ti 16GB | 16 GB | Q8_0 · ~185 tok/s | details |
| RTX 4060 Ti 16GB | 16 GB | Q8_0 · ~119 tok/s | details |
| RX 7800 XT | 16 GB | Q8_0 · ~257 tok/s | details |
| RX 7900 GRE | 16 GB | Q8_0 · ~237 tok/s | details |
| RX 9070 | 16 GB | Q8_0 · ~264 tok/s | details |
| RX 9070 XT | 16 GB | Q8_0 · ~264 tok/s | details |
| RTX 5070 Ti | 16 GB | Q8_0 · ~369 tok/s | details |
| RTX 4070 Ti Super | 16 GB | Q8_0 · ~277 tok/s | details |
| RTX 4080 Super | 16 GB | Q8_0 · ~303 tok/s | details |
| RTX 5080 | 16 GB | Q8_0 · ~395 tok/s | details |
| RTX 4080 | 16 GB | Q8_0 · ~295 tok/s | details |
| RX 7900 XT | 20 GB | F16 · ~289 tok/s | details |
| RX 7900 XTX | 24 GB | F16 · ~347 tok/s | details |
| RTX 3090 | 24 GB | F16 · ~338 tok/s | details |
| RTX 4090 | 24 GB | F16 · ~365 tok/s | details |
| RTX 3090 Ti | 24 GB | F16 · ~365 tok/s | details |
| RTX 5090 | 32 GB | F16 · ~648 tok/s | details |
| DGX Spark | 128 GB | F16 · ~99 tok/s | details |
| Ryzen AI Max+ 395 128GB | 128 GB | F16 · ~93 tok/s | details |
| Mac M1 16GB | 16 GB | No | details |
| Mac M4 16GB | 16 GB | No | details |
| Mac M2 24GB | 24 GB | F16 · ~36 tok/s | details |
| Mac M3 24GB | 24 GB | F16 · ~36 tok/s | details |
| Mac M4 Pro 24GB | 24 GB | F16 · ~99 tok/s | details |
| Mac M4 32GB | 32 GB | F16 · ~43 tok/s | details |
| Mac M1 Pro 32GB | 32 GB | F16 · ~72 tok/s | details |
| Mac M3 Pro 36GB | 36 GB | F16 · ~54 tok/s | details |
| Mac M4 Pro 48GB | 48 GB | F16 · ~99 tok/s | details |
| Mac M4 Pro 64GB | 64 GB | F16 · ~99 tok/s | details |
| Mac M1 Max 64GB | 64 GB | F16 · ~145 tok/s | details |
| Mac M4 Max 64GB | 64 GB | F16 · ~197 tok/s | details |
| Mac M2 Max 96GB | 96 GB | F16 · ~145 tok/s | details |
| Mac M3 Max 128GB | 128 GB | F16 · ~145 tok/s | details |
| Mac M4 Max 128GB | 128 GB | F16 · ~197 tok/s | details |
| Mac M2 Ultra 192GB | 192 GB | F16 · ~289 tok/s | details |
| Mac M3 Ultra 256GB | 256 GB | F16 · ~296 tok/s | details |
| Mac M3 Ultra 512GB | 512 GB | F16 · ~296 tok/s | details |
Questions
How much VRAM do I need for gpt-oss 20B?
About 11.8 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 4.2B of its 21B 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-120b?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.