OpenAI's 117B open model ships its experts in 4-bit MXFP4, so the files are about 63 GB whichever quant you pick. Here is how much VRAM and RAM it needs at 8K to 128K context, which GPUs and machines hold it, and when a 24 GB card plus RAM is enough.
This post contains affiliate links. If you buy through them, Aliteq may earn a commission — at no extra cost to you. Prices verified at publish time.
Share
gpt-oss-120b has 117 billion parameters, which sounds like a server-room number. It isn't quite. OpenAI's model card says it fits "a single 80GB GPU (like NVIDIA H100 or AMD MI300X)," and it is the biggest open model many home builders can realistically run.
I read the model card, the config and the file lists on Hugging Face on 4 October 2026, then ran the numbers through our VRAM engine. I have not benchmarked it for this piece. Everything below is file sizes and arithmetic, with the formula shown.
How much VRAM does gpt-oss-120b need?
About 61 GiB for the weights alone, and about 64 GiB in total with 32K tokens of context. At the full 128K window our engine budgets about 71 GiB. That is the number to plan around, whether the memory is VRAM, unified memory, or VRAM plus system RAM.
The official repo holds 65,248,893,184 bytes of safetensors. That is 60.8 GiB. The model card gives 117B total parameters with 5.1B active per token, and the config sets a 131,072-token maximum context.
Our engine adds the context cache with one formula: 2 (keys and values) x layers x KV heads x head size x tokens x 2 bytes. For gpt-oss-120b that is 2 x 36 x 8 x 64 x 2 bytes, or 72 KiB per token.
gpt-oss-120b total memory by context (GiB, our engine, 4 Oct 2026)
8K tokens
Weights
60.8
Cache
0.6
Total
62.1
32K tokens
Weights
60.8
Cache
2.3
Total
63.8
128K tokens
Weights
60.8
Cache
9.0
Total
70.6
Weights
Cache
Total
8K tokens
60.8
0.6
62.1
32K tokens
60.8
2.3
63.8
128K tokens
60.8
9.0
70.6
Those totals are on the safe side. The config shows that half of the 36 layers use a 128-token sliding window, so their cache stays tiny. Counting only the 18 full-attention layers, 128K tokens costs about 4.5 GiB, for a total near 66 GiB. Plan for the higher figure and enjoy the spare room.
Does it fit? gpt-oss-120b needs about 64 GiB at 32K context
gpt-oss-120b (MXFP4, 32K context) needs ≈64 GB
RTX 4090 (24 GB)24 GBover 40 GB
RTX 5090 (32 GB)32 GBover 32 GB
A100 / H100 (80 GB)80 GBfits
RTX PRO 6000 (96 GB)96 GBfits
A 24 or 32 GB card needs expert offload to system RAM. An 80 GB or 96 GB card holds the whole model.
Will a Q4 or Q2 quant make it smaller?
Barely. OpenAI post-trained the expert weights in MXFP4, a 4-bit format, and those experts are almost all of the model. Community GGUFs keep them as they are. So unsloth's Q2_K is 62.6 GB, Q4_K_M is 62.8 GB and Q8_0 is 63.4 GB. Pick any of them and you need about the same memory.
The config shows what stays at higher precision: attention, the router, the embeddings and the output layer. Those are small. That is why llama.cpp's own ggml-org GGUF (63.4 GB) and Ollama's gpt-oss:120b tag (65 GB) land in the same place.
So skip the hunt for a smaller file. If 64 GiB is out of reach, offload or a smaller model is the fix. For the smaller one, gpt-oss-20b runs in 16 GB, as our Mac mini 16 GB guide works out.
gpt-oss-120b download sizes (read 4 Oct 2026)
Official safetensors (MXFP4 experts)
Size
65.2 GB
Source
OpenAI
Ollama gpt-oss:120b
Size
65 GB
Source
Ollama library
GGUF MXFP4
Size
63.4 GB
Source
ggml-org (llama.cpp)
GGUF Q8_0
Size
63.4 GB
Source
unsloth
GGUF Q4_K_M
Size
62.8 GB
Source
unsloth
GGUF Q2_K
Size
62.6 GB
Source
unsloth
Size
Source
Official safetensors (MXFP4 experts)
65.2 GB
OpenAI
Ollama gpt-oss:120b
65 GB
Ollama library
GGUF MXFP4
63.4 GB
ggml-org (llama.cpp)
GGUF Q8_0
63.4 GB
unsloth
GGUF Q4_K_M
62.8 GB
unsloth
GGUF Q2_K
62.6 GB
unsloth
What GPU do I need for gpt-oss-120b?
One card with 80 GB or more holds it outright: an A100 or H100 80 GB, or a 96 GB RTX PRO 6000. Below that, you combine cards or offload. Two 32 GB RTX 5090s give 64 GB, which is tight at 32K context. A single 24 or 32 GB card needs experts moved to system RAM.
Here is how common setups stack up against the 32K total of about 64 GiB.
Which setups hold gpt-oss-120b (32K context, derived)
How much system RAM do I need to run it on a 24 GB GPU?
About 64 GB of system RAM next to a 24 GB card. The model needs roughly 64 GiB in total at 32K context. With llama.cpp's --n-cpu-moe, attention and some experts stay on the GPU and the rest sit in RAM. 24 GB plus 64 GB covers it with room for your operating system. 32 GB of RAM does not.
This works well because only 5.1B parameters are active per token. Each token touches a small slice of the experts sitting in RAM, so the slowdown is far gentler than offloading a dense model. It is still slower than holding everything in VRAM. We have not measured the speed for this piece. Our --n-cpu-moe explainer shows what the flag does, and the step-by-step offload guide shows how to set it.
On a Mac or a Strix Halo mini PC, the same 64 GiB has to fit in unified memory alongside the operating system. A 64 GB machine is tight. A 96 or 128 GB one is comfortable. Our Strix Halo memory guide and Mac Studio vs NVIDIA build cover those machines.
117B parameters in memory, about 5.1B active per token. The memory is the hard part, not the compute. · Unsplash
Is it cheaper to rent a GPU for gpt-oss-120b?
For occasional use, yes. On 4 October 2026 our tracker showed an 80 GB A100 PCIe at $1.19 an hour and a 96 GB RTX PRO 6000 Max-Q at $1.64 an hour on Runpod, on demand. Either holds the whole model. A few hours a week costs far less than a 96 GB card.
Whether renting beats buying depends on your hours. Our rent vs buy guide does that math, and the cloud GPU tracker shows today's prices.
Referral link
No 80 GB card? Rent one by the hour
An 80 GB A100 PCIe listed at $1.19 an hour and a 96 GB RTX PRO 6000 Max-Q at $1.64 an hour on Runpod on demand (our tracker, 4 Oct 2026). Prices move, so check the live figure before you rent.
Referral link — we may earn a commission at no cost to you. Prices on our compare page are the provider's live figures, cheapest first; this never changes the ranking.
gpt-oss-120b requirements: quick answers
How much VRAM does gpt-oss-120b need?
About 61 GiB for the weights and about 64 GiB in total with 32K tokens of context, by our engine. The full 128K window budgets about 71 GiB. One 80 GB GPU holds it.
What are the gpt-oss-120b system requirements for a home PC?
Either a GPU setup with 64 GB or more of VRAM, or a 24 to 32 GB GPU plus 64 GB of system RAM using expert offload in llama.cpp. A 96 to 128 GB unified-memory machine also works.
Can I run gpt-oss-120b on an RTX 4090 or 3090?
Yes, with offload. Keep attention on the 24 GB card and push most experts to 64 GB of system RAM with --n-cpu-moe. It runs slower than a card that holds the whole model.
Is there a smaller gpt-oss-120b quant?
Not in any useful sense. GGUF files from Q2_K to Q8_0 are 62.6 to 63.4 GB, because the experts stay in their native 4-bit MXFP4 format.
How many parameters are active in gpt-oss-120b?
5.1B per token out of 117B total, per OpenAI's model card. Memory has to hold all 117B; compute only touches the active ones.
The short version: plan for 64 GiB. Get it from one big GPU, several smaller ones, unified memory, or a 24 GB card plus 64 GB of RAM. Try your own setup in the VRAM calculator, and if 64 GB is out of reach, our best MoE models for 12 to 16 GB GPUs has options that fit.
