gpt-oss 20B
gpt-oss 20B needs 13.4 GB of VRAM at Q4_K_M with 8k context, computed from its config.json as checked on 16 October 2026. The smallest card that holds it has 16 GB, and the fastest desktop card that runs it, the RTX 5090, generates at most 601 tokens per second.
- Parameters
- 21B, 3.6B active
- Context
- 131,072 tokens
- Licence
- Apache 2.0
- Released
- August 2025
- Config checked
- 16 October 2026
Open gpt-oss 20B in the LLM checker
Which card size runs it
13.4 GB against each memory size. A model may use 95% of a card, and only what is left over, or missing, is coloured; tight means above 85%, with little room for a longer conversation.
Other quantisations, context lengths or a Q8 KV cache: the VRAM calculator.
Computed 16 October 2026Which GPU to pick
The smallest card here that loads it at Q4_K_M with 8k of context is the RTX 4060 Ti (16 GB): 13.4 GB needed, an estimated 97 tokens per second. For room to work, meaning 32,768 tokens of context while staying under 85% of memory, step up to the RX 7900 XT (20 GB), which needs 14 GB for that and generates at about 268 tokens per second.
At Q8_0, which is near-lossless, the card is the RTX 5090 (32 GB), needing 23.9 GB. On Apple silicon, a Mac mini M4 Pro holds it with 32,768 tokens of context at an estimated 92 tokens per second; prompt processing is slower than on Nvidia, which long documents make obvious. Among unified-memory desktops, the Ryzen AI Max+ 395 holds it at 32,768 tokens in 14 GB, at an estimated 86 tokens per second.
Speed on each card
| GPU | Fits | Est. tokens/s | Feels like | Memory | Bandwidth |
|---|---|---|---|---|---|
| H100 SXM | Yes | ≤ 1,123 | faster than you can read | 80 GB | 3,350 GB/s |
| A100 | Yes | ≤ 684 | faster than you can read | 80 GB | 2,039 GB/s |
| RTX 5090 | Yes | ≤ 601 | faster than you can read | 32 GB | 1,792 GB/s |
| RTX PRO 6000 Blackwell | Yes | ≤ 601 | faster than you can read | 96 GB | 1,792 GB/s |
| RTX PRO 5000 Blackwell | Yes | ≤ 451 | faster than you can read | 48 GB | 1,344 GB/s |
| RTX 4090 | Yes | ≤ 338 | faster than you can read | 24 GB | 1,008 GB/s |
| RTX 5080 | Yes | ≤ 322 | faster than you can read | 16 GB | 960 GB/s |
| RX 7900 XTX | Yes | ≤ 322 | faster than you can read | 24 GB | 960 GB/s |
| RTX 6000 Ada | Yes | ≤ 322 | faster than you can read | 48 GB | 960 GB/s |
| RTX 3090 | Yes | ≤ 314 | faster than you can read | 24 GB | 936 GB/s |
| RTX 5070 Ti | Yes | ≤ 300 | faster than you can read | 16 GB | 896 GB/s |
| RTX PRO 4500 Blackwell | Yes | ≤ 300 | faster than you can read | 32 GB | 896 GB/s |
| L40S | Yes | ≤ 290 | faster than you can read | 48 GB | 864 GB/s |
| RX 7900 XT | Yes | ≤ 268 | faster than you can read | 20 GB | 800 GB/s |
| RTX A6000 | Yes | ≤ 257 | faster than you can read | 48 GB | 768 GB/s |
| RTX 4080 SUPER | Yes | ≤ 247 | faster than you can read | 16 GB | 736 GB/s |
| RTX PRO 4000 Blackwell | Yes | ≤ 225 | faster than you can read | 24 GB | 672 GB/s |
| RX 9070 XT | Yes | ≤ 215 | faster than you can read | 16 GB | 640 GB/s |
| Radeon AI PRO R9700 | Yes | ≤ 215 | faster than you can read | 32 GB | 640 GB/s |
| Arc Pro B70 | Yes | ≤ 204 | faster than you can read | 32 GB | 608 GB/s |
| Arc Pro B60 | Yes | ≤ 153 | faster than you can read | 24 GB | 456 GB/s |
| RTX 5060 Ti | Yes | ≤ 150 | faster than you can read | 16 GB | 448 GB/s |
| RTX PRO 4000 Blackwell SFF | Yes | ≤ 145 | faster than you can read | 24 GB | 432 GB/s |
| RX 9060 XT | Yes | ≤ 107 | faster than you can read | 16 GB | 320 GB/s |
| RTX Spark | Yes | ≤ 101 | faster than you can read | 128 GB (102 for the GPU) | 300 GB/s |
| RTX 4060 Ti | Yes | 97 | faster than you can read | 16 GB | 288 GB/s |
| DGX Spark | Yes | 92 | faster than you can read | 128 GB (126 for the GPU) | 273 GB/s |
| Ryzen AI Max+ 395 | Yes | 86 | faster than you can read | 128 GB (96 for the GPU) | 256 GB/s |
| RTX 3060 | No | - | does not fit | 12 GB | 360 GB/s |
| RTX 4070 | No | - | does not fit | 12 GB | 504 GB/s |
| RTX 5070 | No | - | does not fit | 12 GB | 672 GB/s |
| Arc B580 | No | - | does not fit | 12 GB | 456 GB/s |
Single-stream decode, from memory bandwidth at 70% efficiency. Prompt processing and batching not included. These are ceilings, not measurements: gpt-oss 20B reads only 2.1 GB of weights per token at Q4, so little that the per-token costs the formula leaves out decide much of the real speed. The 25 figures marked ≤ are where a real runtime falls furthest below the number shown. See the speed estimator for other quantisations and Apple hardware, or every card compared if you are choosing hardware rather than a model.
Computed 16 October 2026Memory by quantisation and context
| Quant | 4,096 ctx | 8,192 ctx | 32,768 ctx |
|---|---|---|---|
| Q8_0 | 23.8 GB · fits 32 GB | 23.9 GB · fits 32 GB | 24.5 GB · fits 32 GB |
| Q5_K_M | 16.1 GB · fits 24 GB | 16.2 GB · fits 24 GB | 16.8 GB · fits 24 GB |
| Q4_K_M | 13.3 GB · fits 16 GB | 13.4 GB · fits 16 GB | 14 GB · fits 16 GB |
Weights at this quant: 12.2 GB at Q4. Every extra 1,000 tokens of context adds about 0.025 GB of KV cache at FP16, on top of a fixed 0.003 GB that the sliding-window layers hold once their window is full. Try other settings in the VRAM calculator.
About the model
gpt-oss 20B has 21B parameters with 3.6B active per token (mixture of experts). It has 24 layers with 8 KV heads of dimension 64, alternating full attention with a 128-token sliding window, so only half the layers keep a cache that grows. The smaller of OpenAI's two open-weight models, August 2025. A mixture-of-experts reasoning model that ships already quantised to MXFP4, which is why a 21B model is a 16 GB model.
Text only 131,072-token window Apache 2.0 August 2025 3.6B of 21B active
Reasoning. Always reasons, at low, medium or high effort, set in the system prompt.
Ships in MXFP4; roughly 4.25 bits per expert weight. Choose the 4-bit row.
Native context window: 131,072 tokens. Licence: Apache 2.0, a permissive licence. Weights published in August 2025.
Good at
- A 16 GB card: OpenAI states it runs within 16 GB, and the arithmetic here agrees.
- Speed. Only 3.6B parameters are read per token.
- Tool use and structured output, which it was trained for.
Watch out for
- It expects OpenAI's harmony chat format. Runtimes handle this for you; a hand-rolled prompt template will not.
- There is no higher-precision version to move up to: MXFP4 is how it was trained to ship, so the Q8 row is hypothetical.
- Text only, in a year when most new models read images.
Similar models
The models that need about the same memory as gpt-oss 20B at Q4_K_M and 8k, which makes them the real alternatives on whatever card you have. The context column is what each extra 1,000 tokens costs, and it is where models of the same size differ most.
| Model | Size | Needs | Per 1k context | Window | Released |
|---|---|---|---|---|---|
| gpt-oss 20B | 21B · 3.6B active | 13.4 GB | 0.025 GB | 128k | 2025-08 |
| Phi-4 14B | 14.7B | 11 GB | 0.205 GB | 16k | 2024-12 |
| Devstral Small 2 24B | 24B | 16.3 GB | 0.164 GB | 256k | 2025-12 |
| Mistral Small 3.1 24B | 24B | 16.3 GB | 0.164 GB | 128k | 2025-03 |
| Gemma 4 26B-A4B (MoE) | 25.8B · 3.8B active | 16.4 GB | 0.02 GB | 256k | 2026-03 |
See every model compared, which models fit on 8 to 96 GB, or the cheapest hour of every cloud card.
Questions
- Can I run gpt-oss 20B on a 24 GB card?
- Yes, at Q4_K_M and 8k context it needs about 13.4 GB, which fits a 24 GB card with room.
- How much VRAM does gpt-oss 20B need at Q8?
- About 23.9 GB at 8k context, or 24.5 GB at 32k. Q8 is near-lossless; use it when it fits.
- What is the context length of gpt-oss 20B?
- 131,072 tokens natively. Holding all of it at Q4_K_M takes about 16.4 GB, of which 3.2 GB is context.
- Why does gpt-oss 20B need so little memory for long context?
- Because most of its layers do not keep a cache that grows. It has 24 layers with 8 KV heads of dimension 64, alternating full attention with a 128-token sliding window, so only half the layers keep a cache that grows. At 32k tokens the context costs about 0.8 GB; if all 24 layers kept a conventional full-length cache, the same conversation would cost about 1.6 GB. That saving depends on the runtime implementing the sliding-window cache. llama.cpp and vLLM do; a build that does not will use far more.
- How fast is gpt-oss 20B on an RTX 4090?
- At most about 338 tokens per second at Q4, single stream, which is faster than you can read. That is a ceiling worked out from the card's memory bandwidth, not a measurement, and at a figure this high real runtimes land well below it, because the costs the formula leaves out take a growing share of each token.
- Architecture values from config.json in openai/gpt-oss-20b on Hugging Face, checked 16 October 2026. Every figure on this page is computed from those values; check them before buying hardware for this model.
- Estimates assume a single conversation on a card that is otherwise free. A desktop on the same GPU takes 0.5 to 2 GB.