gpt-oss 120B on a Jetson Orin Nano Super (8GB)

Every quant, against this card

Weight bytes are the size of the real published file wherever one exists — 6 of these6 rows are measured from unsloth/gpt-oss-120b-GGUF, the rest computed from the parameter count. Max context is what the KV cache can grow to in whatever memory the weights leave behind, capped at the 128k tokens this model was trained to address.

What the context actually costs

The KV cache is 2 · layers · kv_heads · head_dim bytes per token per element — 2 · 36 · 8 · 64 · 2 B = 72.0 KiB. It depends on the key/value head count, not the attention-head count and not the parameter count. This model shares 8 KV heads across 64 query heads, which divides the cache by 8 against multi-head attention.

Sliding-window attention. 18 of this model's 36 layers attend to a 128-token window and stop growing there; only the remaining 18 keep scaling with context. That is why the cache figures below flatten out — and why a calculator that ignores the layer pattern over-states this model's long-context footprint several times over.

What running it anyway would cost

Nothing here fits, so the weights would have to be split with part of the model in host RAM. At Q3_K_M that is 52.3 GiB on the host side, and assuming 90 GB/s of host memory bandwidth — a dual-channel DDR5 desktop — the ceiling falls to 27.0 tok/s, against 30.3 tok/s if the same weights were resident — 102 GB/s over the 3.37 GB one token reads, which is the weights the model routes through plus one pass over the cache, against a 62.63 GB weight file. That host bandwidth is an assumption and it is the number to change first if your machine differs; the ratio is the part that generalises.

Where these numbers come from

Questions this pairing answers

The same model on a different card

• GeForce RTX 3060 12GBno fit
• GeForce RTX 5080no fit
• GeForce RTX 5070 Tino fit
• GeForce RTX 4080 SUPERno fit
• GeForce RTX 4070 Ti SUPERno fit
• GeForce RTX 4060 Ti 16GBno fit
• GeForce RTX 4090no fit
• GeForce RTX 3090no fit

• Laguna S 2.1no fit
• Qwen2.5 72B Instructno fit
• DeepSeek-R1-Distill-Llama 70Bno fit
• Llama 3.1 70B Instructno fit
• Llama 3.1 Nemotron 70B Instructno fit
• Llama 3.3 70B Instructno fit
• Mixtral 8x7B Instructno fit
• Hermes 4.3 36Bno fit

Method and limits. Weight bytes are the byte size of the real published file wherever one exists, and the model's exact parameter count times the published llama.cpp bits-per-weight where it does not. That distinction is on every row above and it matters at both ends: a sub-1B model's Q4_K_M file runs a third larger than the nominal figure because k-quants keep its embedding tables at higher precision, and an already-4-bit release cannot be quantized upward at all. The KV cache is 2 · layers · kv_heads · head_dim · bytes per token, summed over layers with each sliding-window layer capped at its window. The speed figure is a roofline bound, not a benchmark: bandwidth divided by bytes read per token, which no runtime exceeds and every runtime falls short of. The usable fraction of card memory is an assumption: 75% here, which is what this lane assumes for unified memory, where the OS and window server share the same pool — the other class assumes 92%, so it is not one number applied to every device. Because nothing here fits, a second assumed input is in play — the 90 GB/s of host memory bandwidth the offload ceiling is priced at, stated above. Those two are the assumed inputs; every other figure is computed from the model config and the card's published specification. Nothing on this page is written by a language model. Architecture from the model's published config, fetched 2026-08-07.