# gpt-oss 120B on a GeForce RTX 4090

No. Even Q3_K_M needs 58.3 GiB against 22.1 GiB usable.

Canonical page: https://makerportal.ai/lab/llm-vram/gpt-oss-120b/rtx-4090
Page title: gpt-oss 120B VRAM on RTX 4090 — does not fit

This markdown document and the HTML page above are rendered from the same solved values at build time, by the same functions. Nothing here is written by a language model and nothing is fetched at request time.

## Key figures

- **Verdict:** Does not fit — 22.1 GiB usable of 24 GB
- **Usable memory:** 22.1 GiB — 92% of the card's 24 GB — an assumption for a dedicated card, not a single fraction applied to every device
- **Weights at FP16:** 60.8 GiB — 116.8B parameters
- **Weights (Q4_K_M):** 58.5 GiB — 60.8 GiB at FP16 · 116.8B params · 6 of 6 quant rows are measured files
- **KV cache:** 72.0 KiB per token at FP16 — 36 layers × 8 KV heads × 64 dimensions
- **Speed ceiling:** 38.9 tok/s — offloaded — nothing fits in device memory
- **Largest context that fits:** none — the weights do not fit
- **Trained context:** 131,072 — the cap the KV-cache figures are held to
- **Card bandwidth:** 1008 GB/s — 24 GB GDDR6X

## Every quant, against this card

Every quantization of gpt-oss 120B against the 22.1 GiB usable on a GeForce RTX 4090. "Measured" rows are the byte size of a real published file; "nominal" rows are the parameter count times the published bits-per-weight.

| Precision | Bits/weight | Weights | Source | Fits | Max context | Ceiling |
|---|---|---|---|---|---|---|
| As released | 4.47 | 60.8 GiB | measured file | short by 38.7 GiB | — | 36.9 tok/s offloaded |
| Q8_0 | 4.34 | 59.0 GiB | measured file | short by 37.0 GiB | — | 38.3 tok/s offloaded |
| Q6_K | 4.33 | 58.9 GiB | measured file | short by 36.9 GiB | — | 38.4 tok/s offloaded |
| Q5_K_M | 4.31 | 58.6 GiB | measured file | short by 36.5 GiB | — | 38.7 tok/s offloaded |
| Q4_K_M | 4.30 | 58.5 GiB | measured file | short by 36.4 GiB | — | 38.8 tok/s offloaded |
| Q3_K_M | 4.29 | 58.3 GiB | measured file | short by 36.2 GiB | — | 38.9 tok/s offloaded |

## Questions this page answers

### How much VRAM does gpt-oss 120B need?

60.8 GiB for the weights at FP16 — 116.8B parameters at two bytes each — and 58.5 GiB at Q4_K_M. The KV cache is on top of that and is not a fixed number: this model spends 72.0 KiB per token of context on its full-attention layers, so 8,192 tokens costs a further 293 MiB. A GeForce RTX 4090 makes 22.1 GiB of its 24 GB available on the assumption below.

### Can a GeForce RTX 4090 run gpt-oss 120B?

No. Even Q3_K_M needs 58.3 GiB against 22.1 GiB usable. Holding the smallest quant here would need a card with about 64 GB. Splitting the model across the PCIe bus is possible and slow — see the offload figure on this page.

### How fast will gpt-oss 120B run on a GeForce RTX 4090?

It cannot run in this card's memory alone, so the speed is set by whatever bus the offloaded part is read across, not by the 1008 GB/s of the card. At an assumed 90 GB/s of host memory bandwidth the ceiling is 38.9 tok/s at Q3_K_M, against 299.2 tok/s if it were resident — 1008 GB/s over the 3.37 GB one token reads at Q3_K_M, priced at the 8k reference context. That is the weights the model routes through plus one pass over the cache, against a 62.63 GB weight file.

### Why does the context length change how much memory gpt-oss 120B needs?

Because the KV cache holds one key and one value vector per token, per layer, for the whole conversation, and it is allocated separately from the weights. This model has 36 layers and 8 key/value heads of 64 dimensions, shared across 64 query heads — grouped-query attention, which divides the cache by 8. That works out at 72.0 KiB per token, except that 18 of the 36 layers use a 128-token sliding window and stop growing there. Parameter count tells you nothing about this number.

## 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: 92% here, which is what this lane assumes for a dedicated card — the other class assumes 75%, 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.

## Related tool

[LLM VRAM & KV-Cache Footprint Calculator](https://makerportal.ai/lab/llm-vram-kvcache-calculator) — Interactive VRAM planner — change the model geometry, quantization, context length and card capacity and read the weight bytes, KV-cache bytes and the bandwidth ceiling on decode speed.

---

Source: MakerPortal — https://makerportal.ai/lab/llm-vram/gpt-oss-120b/rtx-4090. Free to quote and cite with attribution and a link to the canonical page.