Hardware guide · October 2026
What GPU do you need to run gpt-oss-120b?
gpt-oss-120b is a 117B (5.1B active) mixture-of-experts model. Here is exactly how much memory it needs and which GPUs and Macs run it well.
The short answer
To run gpt-oss-120b at MXFP4 with 8K tokens of context you need about 68GB of VRAM (65.4GB of weights plus KV cache and overhead). The cheapest way in is a 72GB card like the RTX PRO 5000 72GB (~$9,100). For the best speed per dollar, get the RTX PRO 6000 (~170 tokens/sec). On a Mac or other unified-memory system, the Ryzen AI Max+ 395 runs it at ~35 tokens/sec.
- VRAM at MXFP4
- ~68GB
- GPUs that run it
- 10
- Parameters
- 117B
- CPU only (est.)
- ~15 tok/s
Our recommendations
The best GPUs for gpt-oss-120b
Minimum
RTX PRO 5000 72GB72GB · ~$9,100
- Quant
- MXFP4
- Speed (est.)
- ~145 tok/s
The smallest card that runs it at MXFP4 with 8K context.
Best value
RTX PRO 600096GB · ~$16,000
- Quant
- MXFP4
- Speed (est.)
- ~170 tok/s
The most tokens per second per dollar among cards with room to spare.
Multi-GPU
3× RTX 309072GB total · ~$3,150 used
- Quant
- MXFP4
- Speed (est.)
- ~66 tok/s
Split across cards with llama.cpp, vLLM, or ExLlama. Needs a board and PSU that can feed every card.
Faster multi-GPU
3× RTX 409072GB total · ~$7,500 used
- Quant
- MXFP4
- Speed (est.)
- ~71 tok/s
Split across cards with llama.cpp, vLLM, or ExLlama. Needs a board and PSU that can feed every card.
Mac / unified memory
Ryzen AI Max+ 395128GB unified · ~$3,300
- Quant
- MXFP4
- Speed (est.)
- ~35 tok/s
Lots of memory per dollar and silent, but slower than a GPU with the same capacity.
We may earn a commission when you buy through links on this page. It never changes what we recommend. How we make money
Memory
gpt-oss-120b VRAM requirements
gpt-oss-120b is released in MXFP4, so its memory use depends mostly on context length.
| Quantization | Quality | Weights | Total @ 8K ctx | Total @ 32K ctx |
|---|---|---|---|---|
| MXFP4 | NativeDefault | 65 GB | 68 GB | 69 GB |
Context
How context length changes the memory you need
Every token in the context window is stored in the KV cache. At MXFP4, with a 16-bit KV cache, gpt-oss-120b needs:
| Context | KV cache | Total memory | With 8-bit KV |
|---|---|---|---|
| 4K tokens | 0.2 GB | 68.1 GB | 68.0 GB |
| 8K tokens | 0.3 GB | 68.3 GB | 68.1 GB |
| 16K tokens | 0.6 GB | 68.6 GB | 68.3 GB |
| 32K tokens | 1.2 GB | 69.2 GB | 68.6 GB |
| 64K tokens | 2.4 GB | 70.4 GB | 69.2 GB |
| 128K tokens | 4.8 GB | 72.8 GB | 70.4 GB |
36 layers alternate between a 128-token sliding window and full attention, so only the 18 full-attention layers grow with context: 2 x 18 x 8 KV heads x 64 dims x 2 bytes = 36,864 bytes (36 KiB) per token, about 4.8GB at the full 128K. The sliding layers add a negligible fixed buffer.
Compatibility
Every GPU that runs gpt-oss-120b
Each GPU at the highest-quality quantization that fits with 8K context, fastest first. Speeds are single-user estimates.
| GPU | Memory | Quant | Speed | Fit | Price |
|---|---|---|---|---|---|
| RTX PRO 6000 | 96GB | MXFP4 | ~170 tok/s | Runs well | ~$16,000 |
| M5 Ultra 512GB | 512GB unified | MXFP4 | ~93 tok/s | Runs well | — |
| M5 Ultra 256GB | 256GB unified | MXFP4 | ~93 tok/s | Runs well | ~$9,499 |
| M3 Ultra 512GB | 512GB unified | MXFP4 | ~70 tok/s | Runs well | $9,499 |
| M5 Max 128GB | 128GB unified | MXFP4 | ~64 tok/s | Runs well | ~$5,100 |
| M4 Max 128GB | 128GB unified | MXFP4 | ~58 tok/s | Runs well | $3,499 |
| Ryzen AI Max+ PRO 495 | 192GB unified | MXFP4 | ~37 tok/s | Runs well | ~$6,800 |
| DGX Spark | 128GB unified | MXFP4 | ~37 tok/s | Runs well | ~$6,950 |
| Ryzen AI Max+ 395 | 128GB unified | MXFP4 | ~35 tok/s | Runs well | ~$3,300 |
| RTX PRO 5000 72GB | 72GB | MXFP4 | ~145 tok/s | Fits, tight | ~$9,100 |
With CPU offload (partial offload, assuming 64GB of system RAM): RTX 6000 Ada (~37 tok/s), RTX A6000 (~35 tok/s), RTX 5090 (~25 tok/s), V100 32GB (~24 tok/s), R9700 (~23 tok/s), MI50 32GB (~22 tok/s), Arc Pro B70 (~21 tok/s), RTX 4090 (~21 tok/s) and 4 more.
Tips
Getting the most out of gpt-oss-120b
- Budget memory for context. The KV cache grows with every token. Quantizing it to 8-bit roughly halves that cost with little quality loss.
- Offload experts, not layers. For mixture-of-experts models, llama.cpp's
--n-cpu-moe(or-ottensor overrides) keeps attention on the GPU and moves expert weights to system RAM. That keeps speed usable when the model does not fit in VRAM. - Pick the right runtime. Ollama and LM Studio are easiest. llama.cpp gives the most control. vLLM or SGLang are best when several people or agents share one machine.
Want to try different quantizations, context lengths, or GPU counts? Use our VRAM calculator, or read the gpt-oss-120b overview for setup commands and benchmarks.
Frequently asked questions
How much VRAM do I need to run gpt-oss-120b?
About 68GB at MXFP4 with an 8K context window. gpt-oss-120b ships in MXFP4, so there are no smaller official quantizations. Longer contexts add 0.04GB per 1K tokens with a 16-bit KV cache.
Can I run gpt-oss-120b on a 24GB GPU?
Only with partial CPU offload. On a 24GB card like the RTX 3090 or RTX 4090, about 68% of the weights would sit in system RAM and generation drops to about 21 tokens/sec with CPU offload.
Can I run gpt-oss-120b on a Mac?
Yes. With 128GB of unified memory or more (for example the Ryzen AI Max+ 395), macOS can give the GPU enough memory to run it at MXFP4, at about ~35 tokens/sec. MLX builds are usually fastest on Apple silicon.
Can I run gpt-oss-120b on CPU only, without a GPU?
Yes, if you have at least 80GB of system RAM. On a desktop with dual-channel DDR5, expect roughly 15 tokens/sec, which is better than you might expect because only the active experts are read for each token. A GPU is recommended for interactive use.