gpt-oss-20b Hardware Requirements
OpenAI's open-weight reasoner, shipped natively in 4-bit (MXFP4) so the whole thing fits in ~13GB. Adjustable reasoning effort and real tool-use training make it a superb local agent brain on 16GB cards.
VRAM needed (Q4, 8k context)
12.7 GB
Cheapest GPU that runs it: Arc A770 ($300–350 street)
Check Price on AmazonUpdated July 2026. Estimates — see methodology below.
VRAM by Quantization
Weights + KV cache at 8k context + 1.2GB system overhead. Q4_K_M is the community default — quality loss is negligible for most use.
| Quantization | Weights | KV cache (8k) | Total VRAM | Cheapest GPU that fits |
|---|---|---|---|---|
| Q4_K_M Recommended — near-lossless for most use, half the size of Q8 | 11.2 GB | 0.4 GB | 12.7 GB | Arc A770 (from $300) |
| Q5_K_M Slightly higher quality than Q4 for ~18% more VRAM | 15.0 GB | 0.4 GB | 16.5 GB | RX 7900 XT (~$588 used) |
| Q8_0 Effectively lossless — use if you have VRAM to spare | 22.3 GB | 0.4 GB | 23.9 GB | RX 7900 XTX (~$838 used) |
| FP16 Full precision — only for fine-tuning or maximum fidelity | 42.0 GB | 0.4 GB | 43.6 GB | Multi-GPU / Mac territory |
* gpt-oss-20b ships natively quantized (~4.25 bits/weight) — the Q4 row reflects its actual release format.
Longer context costs VRAM
KV cache grows linearly with context: 8k → 0.4 GB · 32k → 1.5 GB · 128k → 6.0 GB. If you plan to feed whole documents or codebases, size your GPU for the context you actually need, not just the weights.
Why gpt-oss-20b is fast but VRAM-hungry
gpt-oss-20b is a Mixture-of-Experts model: all 21B parameters must sit in memory, but each token only activates 3.6B of them. Memory capacity requirements are those of a 21B model, while speed is that of a 3.6B model — which is why MoE models feel so fast when they fit, and why Macs with large unified memory punch above their weight running them.
Best GPUs for gpt-oss-20b
The cheapest way to run gpt-oss-20b well. Expect instant-feeling responses at ~88 tokens/sec.
The fastest single-GPU experience for gpt-oss-20b. Expect instant-feeling responses at ~281 tokens/sec.
GPU Compatibility (Q4, 8k context)
Every GPU in our database, scored against gpt-oss-20b. Speed is estimated decode rate — memory-bandwidth-bound, so VRAM and bandwidth matter more than shader count.
| GPU | VRAM | Verdict | Est. speed | Price | |
|---|---|---|---|---|---|
| Arc A770 | 16 GB | Runs great | ~88 tok/sInstant-feeling | from $300 | Check price |
| RX 6800 XT | 16 GB | Runs great | ~80 tok/sInstant-feeling | ~$438 used | Check price |
| RX 7800 XT | 16 GB | Runs great | ~98 tok/sInstant-feeling | ~$488 used | Check price |
| RX 7900 XT | 20 GB | Runs great | ~125 tok/sInstant-feeling | ~$588 used | Check price |
| RTX 5060 Ti | 16 GB | Runs great | ~70 tok/sInstant-feeling | from $550 | Check price |
| RX 9070 | 16 GB | Runs great | ~100 tok/sInstant-feeling | from $575 | Check price |
| RX 9070 XT | 16 GB | Runs great | ~100 tok/sInstant-feeling | from $600 | Check price |
| RTX 4070 Ti SUPER | 16 GB | Runs great | ~105 tok/sInstant-feeling | ~$750 used | Check price |
| RX 7900 XTX | 24 GB | Runs great | ~151 tok/sInstant-feeling | ~$838 used | Check price |
| RTX 4080 SUPER | 16 GB | Runs great | ~115 tok/sInstant-feeling | ~$900 used | Check price |
| RTX 5070 Ti | 16 GB | Runs great | ~141 tok/sInstant-feeling | from $900 | Check price |
| RTX 3090 | 24 GB | Runs great | ~147 tok/sInstant-feeling | ~$1,150 used | Check price |
| RTX 5080 | 16 GB | Runs great | ~151 tok/sInstant-feeling | from $1,250 | Check price |
| RTX 4090 | 24 GB | Runs great | ~158 tok/sInstant-feeling | ~$2,375 used | Check price |
| RTX 5090 | 32 GB | Runs great | ~281 tok/sInstant-feeling | from $2,800 | Check price |
| RTX 3060 | 12 GB | Partial offload | — | ~$238 used | |
| Arc B570 | 10 GB | Partial offload | — | from $225 | |
| RTX 4060 | 8 GB | Partial offload | — | ~$275 used | |
| RX 7600 | 8 GB | Partial offload | — | from $250 | |
| Arc B580 | 12 GB | Partial offload | — | from $250 | |
| RX 6700 XT | 12 GB | Partial offload | — | ~$315 used | |
| RTX 4060 Ti | 8 GB | Partial offload | — | ~$338 used | |
| RTX 3070 | 8 GB | Partial offload | — | ~$338 used | |
| RTX 5060 | 8 GB | Partial offload | — | from $325 | |
| RX 7700 XT | 12 GB | Partial offload | — | ~$415 used | |
| RTX 3080 | 10 GB | Partial offload | — | ~$463 used | |
| RTX 4070 | 12 GB | Partial offload | — | ~$500 used | |
| RTX 4070 SUPER | 12 GB | Partial offload | — | ~$563 used | |
| RTX 5070 | 12 GB | Partial offload | — | from $600 |
Run it in one command
With Ollama installed, this pulls the default quant and starts chatting:
Frequently Asked Questions
How much VRAM do I need to run gpt-oss-20b?+
At the recommended Q4_K_M quantization with 8k context, gpt-oss-20b needs roughly 12.7GB of VRAM (11.2GB weights + KV cache + overhead). Q8 needs about 23.9GB and full FP16 about 43.6GB.
What is the cheapest GPU that runs gpt-oss-20b?+
Intel Arc A770 (16GB, $300–350 street) is the cheapest current GPU in our database that runs gpt-oss-20b fully in VRAM at an estimated ~88 tokens/sec.
Can I run gpt-oss-20b on an RTX 3060?+
Only partially — the RTX 3060 12GB can offload some layers to system RAM, but expect a large speed penalty.
Can I run gpt-oss-20b on a Mac?+
Yes, if the Mac has enough unified memory: budget roughly 12.7GB of RAM for the Q4 version (plus what macOS itself uses). Apple Silicon runs GGUF models well via Ollama or LM Studio, and MoE models like this one are particularly Mac-friendly — only 3.6B parameters are active per token, so memory bandwidth goes further.
Can I run gpt-oss-20b on CPU only?+
Sort of. Because only 3.6B of 21B parameters are active per token, CPU inference is more viable than for dense models this size — expect single-digit tokens/sec with fast DDR5. A GPU is still dramatically better.
Is gpt-oss-20b free for commercial use?+
Yes. gpt-oss-20b is released under the Apache 2.0, which permits commercial use.
Related Models
How we calculate these numbers
VRAM = model weights (parameters × bits per weight ÷ 8) + KV cache (architecture-specific bytes per token × context length) + ~1.2GB runtime overhead. Speed estimates assume decode is memory-bandwidth-bound at ~50% utilization (lower for MoE models, which pay routing overhead), matching typical llama.cpp performance on consumer cards; real results vary with runtime, drivers, and settings. Quant sizes reflect GGUF K-quants, which keep some layers at higher precision. Figures are estimates for planning, not guarantees — when in doubt, buy more VRAM than you need today. Prices shown are launch MSRP; mid-2026 street prices often run well above MSRP due to the ongoing memory shortage, and used 24GB cards are holding their value unusually well.