gpt-oss-20b hardware requirements
What is gpt-oss-20b?
gpt-oss-20b is an open-weight reasoning model that OpenAI released in August 2025. Its model card aims it at local and lower-latency use, with adjustable reasoning effort. It is trained to call functions, browse the web and run Python code, but the app running it has to supply those tools. It is a mixture-of-experts model with 3.6B of its 21B parameters active per token, released under the Apache 2.0 license.
What to know before running gpt-oss-20b
gpt-oss-20b ships in a single format, MXFP4, so there is no smaller quant to fall back on: the file is 12.1 GB and it needs 12.3 GB with an 8k context. That puts the line at 16GB of VRAM. On a 16GB card it runs fully on the GPU and runs great; see our page for gpt-oss-20b on the GeForce RTX 5060 Ti 16GB. It just misses on 12GB cards and is well over 8GB, but as a mixture-of-experts model it can keep the experts that do not fit in system RAM, and CanRun still expects it to run well on the GeForce RTX 3060 12GB and the GeForce RTX 4060. Those two are theoretical estimates (±30%).
It always reasons before answering. The model card lets you set the reasoning effort to low, medium or high, but CanRun counts it as a thinking model either way and holds it to the 1.5× stricter speed bar. On an 8GB card such as the GeForce RTX 4060 Laptop, where part of its experts sit in system RAM, it still clears that bar and runs well (a theoretical estimate, ±30%). On a Mac, 24GB of unified memory is enough; the Apple M4 Pro runs it great.
About the model
- Parameters
- 20.9B total · 3.6B active (MoE)
- Attention
- Sliding-window attention
- Max context
- 131,072 tokens
- Released
- 2025-08-05
- Model card
- gpt-oss-20b on Hugging Face
Other sizes in this series
| Model | Parameters | Baseline quant | Memory needed (8k) | GeForce RTX 4060 | GeForce RTX 5060 Ti 16GB | GeForce RTX 3090 |
|---|---|---|---|---|---|---|
| gpt-oss-20bthis model | 20.9B total · 3.6B active (MoE) | MXFP4 | 12.3 GB | Runs well | Runs great | Runs great |
| gpt-oss-120b | 116.8B total · 5.1B active (MoE) | MXFP4 | 63.6 GB | Won't run | Won't run | Won't run |
GGUF files
| Quant | File size | Quality | Published by |
|---|---|---|---|
| MXFP4baseline | 12.11 GB | Near-lossless | llama.cpp team · ggml-org |
Memory needed by quant and context
| Quant | 4k | 8k | 16k | 32k | 64k | 128k |
|---|---|---|---|---|---|---|
| MXFP4 | 12.1 | 12.3 | 12.5 | 13.1 | 14.2 | 16.4 |
Which GPUs can run gpt-oss-20b?
How to read the verdicts
- Runs great
- Fully on the GPU at 20 tok/s or more
- Runs well
- Fully on the GPU at 8–20 tok/s; MoE experts in system RAM at 20 tok/s or more; or at least 90% on the GPU at 8 tok/s or more
- Runs slowly
- 2–8 tok/s; CPU-only; less than 90% on the GPU; or MoE experts in system RAM below 20 tok/s
- Won't run
- Does not fit, or under 2 tok/s
Filter this list by budget, type and context in the calculator
Scroll sideways to see every column.
| Hardware | Verdict | Speed | Quant | Memory | Runs as | Context | Notes | Price |
|---|---|---|---|---|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Max-Q | Runs great | 373.9 tok/s8k estimate 338.9 tok/s (271.1–406.7, calibrated estimate ±20%)measured (1 run, 512 context) | MXFP4 | 12.9 / 96.0 GB | Full GPU | up to 128k | — | — |
| NVIDIA RTX PRO 6000 Blackwell | Runs great | est. 338.9 tok/s271.1–406.7calibrated estimate ±20% | MXFP4 | 12.9 / 96.0 GB | Full GPU | up to 128k | — | — |
| GeForce RTX 5090 | 282.3 tok/s8k estimate 338.9 tok/s (271.1–406.7, calibrated estimate ±20%)measured (1 run, 2k context) | MXFP4 | 12.9 / 32.0 GB | Full GPU | up to 128k | — | $4,200 street (as of 2026-08-09) | |
| GeForce RTX 4090 | 225.2 tok/s8k estimate 214.1 tok/s (171.3–256.9, calibrated estimate ±20%)measured (1 run, 2k context) | MXFP4 | 12.9 / 24.0 GB | Full GPU | up to 128k | — | $2,755 street (as of 2026-08-09) | |
| GeForce RTX 3090 Ti | Runs great | est. 214.1 tok/s171.3–256.9calibrated estimate ±20% | MXFP4 | 12.9 / 24.0 GB | Full GPU | up to 128k | — | $1,999 launch MSRP |
| GeForce RTX 5080 | Runs great | 204.9 tok/s8k estimate 205.5 tok/s (164.4–246.6, calibrated estimate ±20%)measured (1 run, 2k context) | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | $1,256 street (as of 2026-08-09) |
| GeForce RTX 5080 Laptop | Runs great | est. 193.7 tok/s155.0–232.5calibrated estimate ±20% | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | — |
| GeForce RTX 5090 Laptop | Runs great | est. 193.7 tok/s155.0–232.5calibrated estimate ±20% | MXFP4 | 12.9 / 24.0 GB | Full GPU | up to 128k | — | — |
| GeForce RTX 5070 Ti | Runs great | 189.5 tok/s8k estimate 193.7 tok/s (155.0–232.5, calibrated estimate ±20%)measured (1 run, 2k context) | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | $949 street (as of 2026-08-09) |
| Radeon RX 7900 XTX | Runs great | est. 189.0 tok/s151.2–226.8calibrated estimate ±20% | MXFP4 | 12.9 / 24.0 GB | Full GPU | up to 128k | — | $999 launch MSRP |
| Apple M5 Ultra · 96 GB | Runs great | est. 187.4 tok/s149.9–224.9calibrated estimate ±20% | MXFP4 | 12.9 / 67.2 GB | Unified memory | up to 128k | — | — |
| GeForce RTX 4080 Super | Runs great | 186.4 tok/s8k estimate 163.3 tok/s (130.7–196.0, calibrated estimate ±20%)measured (1 run, 2k context) | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | $999 launch MSRP |
| Radeon RX 7900 XT | Runs great | est. 165.3 tok/s132.2–198.3calibrated estimate ±20% | MXFP4 | 12.9 / 20.0 GB | Full GPU | up to 128k | — | $899 launch MSRP |
| GeForce RTX 3090 | 162.0 tok/s8k estimate 201.1 tok/s (160.9–241.3, calibrated estimate ±20%)measured (1 run, 2k context) | MXFP4 | 12.9 / 24.0 GB | Full GPU | up to 128k | — | $1,050 street (as of 2026-08-09) | |
| GeForce RTX 4080 | Runs great | est. 159.6 tok/s127.7–191.5calibrated estimate ±20% | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | $1,199 launch MSRP |
| GeForce RTX 4070 Ti Super | Runs great | est. 150.7 tok/s120.6–180.8calibrated estimate ±20% | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | $799 launch MSRP |
| Apple M5 Max (40-core GPU) · 48 GB | Runs great | est. 145.2 tok/s116.2–174.3calibrated estimate ±20% | MXFP4 | 12.9 / 33.6 GB | Unified memory | up to 128k | — | — |
| Radeon RX 9070 | Runs great | est. 139.9 tok/s111.9–167.9calibrated estimate ±20% | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | $639 street (as of 2026-08-09) |
| Radeon RX 9070 XT | Runs great | est. 139.9 tok/s111.9–167.9calibrated estimate ±20% | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | $689 street (as of 2026-08-09) |
| Radeon RX 7800 XT | Runs great | est. 136.3 tok/s109.0–163.6calibrated estimate ±20% | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | $499 launch MSRP |
| Apple M2 Ultra · 64 GB | Runs great | est. 134.1 tok/s107.3–160.9calibrated estimate ±20% | MXFP4 | 12.9 / 44.8 GB | Unified memory | up to 128k | — | — |
| GeForce RTX 4090 Laptop | Runs great | est. 131.3 tok/s105.0–157.5calibrated estimate ±20% | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | — |
| Apple M4 Max (40-core GPU) · 48 GB | Runs great | est. 131.1 tok/s104.9–157.4calibrated estimate ±20% | MXFP4 | 12.9 / 33.6 GB | Unified memory | up to 128k | — | — |
| Apple M3 Ultra · 96 GB | Runs great | 115.5 tok/s8k estimate 136.7 tok/s (109.4–164.1, calibrated estimate ±20%)measured (1 run, 2k context) | MXFP4 | 12.9 / 67.2 GB | Unified memory | up to 128k | — | — |
| Apple M5 Max (32-core GPU) · 36 GB | Runs great | est. 112.7 tok/s90.1–135.2calibrated estimate ±20% | MXFP4 | 12.9 / 25.2 GB | Unified memory | up to 128k | — | — |
| GeForce RTX 5060 Ti 16GB | 111.6 tok/s8k estimate 104.3 tok/s (83.5–125.2, calibrated estimate ±20%)measured (1 run, 2k context) | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | $569 street (as of 2026-08-09) | |
| Apple M4 Max (32-core GPU) · 36 GB | Runs great | 92.4 tok/s8k estimate 101.6 tok/s (81.3–121.9, calibrated estimate ±20%)measured (1 run, 2k context) | MXFP4 | 12.9 / 25.2 GB | Unified memory | up to 128k | — | — |
| Radeon RX 9060 XT 16GB | Runs great | est. 78.0 tok/s62.4–93.6calibrated estimate ±20% | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | $449 street (as of 2026-08-09) |
| Apple M5 Pro · 24 GB | Runs great | est. 77.9 tok/s62.4–93.5calibrated estimate ±20% | MXFP4 | 12.9 / 16.8 GB | Unified memory | up to 64k | — | — |
| NVIDIA DGX Spark · 128 GB | Runs great | est. 71.7 tok/s57.4–86.0calibrated estimate ±20% | MXFP4 | 12.9 / 116.9 GB | Unified memory | up to 128k | — | $3,999 launch MSRP (128 GB) |
| Apple M4 Pro · 24 GB | Runs greatgpt-oss-20b verdict on the Apple M4 Pro | est. 69.9 tok/s55.9–83.9calibrated estimate ±20% | MXFP4 | 12.9 / 16.8 GB | Unified memory | up to 64k | — | — |
| GeForce RTX 4060 Ti 16GB | Runs great | est. 69.0 tok/s55.2–82.8calibrated estimate ±20% | MXFP4 | 12.9 / 16.0 GB | Full GPU | up to 64k | — | $499 launch MSRP |
| Ryzen AI Max+ 395 (Strix Halo) · 32 GB | Runs great | est. 68.6 tok/s54.8–82.3calibrated estimate ±20% | MXFP4 | 12.9 / 22.4 GB | Unified memory | up to 128k | — | $1,999 launch MSRP (128 GB) |
| Apple M6 · 24 GB | Runs great | est. 44.8 tok/s35.8–53.7calibrated estimate ±20% | MXFP4 | 12.9 / 16.8 GB | Unified memory | up to 64k | — | — |
| Apple M5 · 24 GB | Runs great | est. 40.5 tok/s32.4–48.6calibrated estimate ±20% | MXFP4 | 12.9 / 16.8 GB | Unified memory | up to 32k | — | — |
| Apple M4 · 24 GB | Runs greatgpt-oss-20b verdict on the Apple M4 | est. 31.9 tok/s25.5–38.3calibrated estimate ±20% | MXFP4 | 12.9 / 16.8 GB | Unified memory | up to 8k | — | — |
| GeForce RTX 3080 12GB | Runs well | est. 105.4 tok/s73.8–137.0theoretical estimate ±30% | MXFP4 | 11.2 / 12.0 GB + 2.3 GB RAM | MoE experts in RAM | up to 128k |
| $799 launch MSRP |
| GeForce RTX 5070 | est. 91.6 tok/s64.1–119.1theoretical estimate ±30% | MXFP4 | 11.2 / 12.0 GB + 2.3 GB RAM | MoE experts in RAM | up to 128k |
| $629 street (as of 2026-08-09) | |
| GeForce RTX 5070 Ti Laptop | Runs well | est. 91.6 tok/s64.1–119.1theoretical estimate ±30% | MXFP4 | 11.2 / 12.0 GB + 2.3 GB RAM | MoE experts in RAM | up to 128k |
| — |
| GeForce RTX 4070 | est. 78.5 tok/s55.0–102.1theoretical estimate ±30% | MXFP4 | 11.2 / 12.0 GB + 2.3 GB RAM | MoE experts in RAM | up to 128k |
| $599 launch MSRP | |
| GeForce RTX 4070 Super | Runs well | est. 78.5 tok/s55.0–102.1theoretical estimate ±30% | MXFP4 | 11.2 / 12.0 GB + 2.3 GB RAM | MoE experts in RAM | up to 128k |
| $599 launch MSRP |
| GeForce RTX 2080 Ti | Runs well | est. 73.9 tok/s51.7–96.0theoretical estimate ±30% | MXFP4 | 10.3 / 11.0 GB + 3.2 GB RAM | MoE experts in RAM | up to 128k |
| $999 launch MSRP |
| GeForce RTX 4080 Laptop | Runs well | est. 71.7 tok/s50.2–93.2theoretical estimate ±30% | MXFP4 | 11.2 / 12.0 GB + 2.3 GB RAM | MoE experts in RAM | up to 128k |
| — |
| GeForce RTX 3080 10GB | Runs well | est. 68.4 tok/s47.9–88.9theoretical estimate ±30% | MXFP4 | 9.5 / 10.0 GB + 4.0 GB RAM | MoE experts in RAM | up to 128k |
| $699 launch MSRP |
| GeForce GTX 1080 Ti | Runs well | est. 66.3 tok/s46.4–86.2theoretical estimate ±30% | MXFP4 | 10.3 / 11.0 GB + 3.2 GB RAM | MoE experts in RAM | up to 64k |
| $699 launch MSRP |
| GeForce RTX 3060 12GB | est. 63.9 tok/s44.8–83.1theoretical estimate ±30% | MXFP4 | 11.2 / 12.0 GB + 2.3 GB RAM | MoE experts in RAM | up to 64k |
| $250 street (as of 2026-08-09) | |
| Intel Arc B580 | Runs well | est. 63.5 tok/s44.5–82.6theoretical estimate ±30% | MXFP4 | 11.2 / 12.0 GB + 2.3 GB RAM | MoE experts in RAM | up to 64k |
| $290 street (as of 2026-08-09) |
| GeForce RTX 3070 | Runs well | est. 43.9 tok/s30.7–57.1theoretical estimate ±30% | MXFP4 | 7.3 / 8.0 GB + 6.1 GB RAM | MoE experts in RAM | up to 64k |
| $499 launch MSRP |
| GeForce RTX 5060 | Runs well | est. 43.9 tok/s30.7–57.1theoretical estimate ±30% | MXFP4 | 7.3 / 8.0 GB + 6.1 GB RAM | MoE experts in RAM | up to 64k |
| $339 street (as of 2026-08-09) |
| GeForce RTX 5060 Ti 8GB | Runs well | est. 43.9 tok/s30.7–57.1theoretical estimate ±30% | MXFP4 | 7.3 / 8.0 GB + 6.1 GB RAM | MoE experts in RAM | up to 64k |
| $429 street (as of 2026-08-09) |
| GeForce RTX 4060 Ti 8GB | Runs well | est. 37.8 tok/s26.5–49.1theoretical estimate ±30% | MXFP4 | 7.3 / 8.0 GB + 6.1 GB RAM | MoE experts in RAM | up to 32k |
| $399 launch MSRP |
| GeForce RTX 4060 | est. 36.9 tok/s25.9–48.0theoretical estimate ±30% | MXFP4 | 7.3 / 8.0 GB + 6.1 GB RAM | MoE experts in RAM | up to 32k |
| $299 launch MSRP | |
| GeForce RTX 4060 Laptop | Runs well | est. 36.0 tok/s25.2–46.8theoretical estimate ±30% | MXFP4 | 7.3 / 8.0 GB + 6.1 GB RAM | MoE experts in RAM | up to 32k |
| — |
| GeForce RTX 4070 Laptop | Runs well | est. 36.0 tok/s25.2–46.8theoretical estimate ±30% | MXFP4 | 7.3 / 8.0 GB + 6.1 GB RAM | MoE experts in RAM | up to 32k |
| — |
| GeForce RTX 3050 8GB | Runs well | est. 34.0 tok/s23.8–44.2theoretical estimate ±30% | MXFP4 | 7.3 / 8.0 GB + 6.1 GB RAM | MoE experts in RAM | up to 16k |
| $249 launch MSRP |
| DDR5-6000 dual-channel CPU · 32 GB | Runs slowly | est. 18.0 tok/s14.4–21.7calibrated estimate ±20% | MXFP4 | 12.3 GB RAM | CPU only | up to 128k |
| — |
| DDR5-5600 dual-channel CPU · 32 GB | Runs slowly | est. 16.8 tok/s13.5–20.2calibrated estimate ±20% | MXFP4 | 12.3 GB RAM | CPU only | up to 128k |
| — |
| DDR4-3200 dual-channel CPU · 32 GB | Runs slowly | est. 9.6 tok/s7.7–11.6calibrated estimate ±20% | MXFP4 | 12.3 GB RAM | CPU only | up to 128k | — | — |
Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.
Cheapest GPUs that run it
- Runs greatRadeon RX 9060 XT 16GB$449 street (as of 2026-08-09), est. 78.0 tok/s (62.4–93.6, calibrated estimate ±20%).
- Runs wellGeForce RTX 3060 12GB$250 street (as of 2026-08-09), est. 63.9 tok/s (44.8–83.1, theoretical estimate ±30%).
Measured results for gpt-oss-20b
Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.
| Hardware | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Max-Q | MXFP4 | llama.cpp | 512 | 17,385.0 | 373.9 | CUDA, -fa 1 | github.com | 2026-06-24 |
| GeForce RTX 5090 | MXFP4 | llama.cpp | 2k | 9,841.0 | 282.3 | — | github.com | 2025-08-20 |
| GeForce RTX 4090 | MXFP4 | llama.cpp | 2k | 8,078.0 | 225.2 | — | github.com | 2025-08-20 |
| GeForce RTX 5080 | MXFP4 | llama.cpp | 2k | 7,469.0 | 204.9 | — | github.com | 2025-08-20 |
| GeForce RTX 5070 Ti | MXFP4 | llama.cpp | 2k | 6,326.0 | 189.5 | — | github.com | 2025-08-20 |
| GeForce RTX 4080 Super | MXFP4 | llama.cpp | 2k | 8,111.0 | 186.4 | — | github.com | 2025-08-29 |
| GeForce RTX 3090 | MXFP4 | llama.cpp | 2k | 5,143.0 | 162.0 | — | github.com | 2025-08-20 |
| GeForce RTX 5060 Ti 16GB | MXFP4 | llama.cpp | 2k | 3,821.0 | 111.6 | — | github.com | 2025-08-20 |
| Radeon RX 7900 XT | BF16 | llama.cpp | 2k | 4,252.0 | 101.9 | ROCm 6.4.1 | github.com | 2025-08-18 |
| Apple M3 Ultra | MXFP4 | llama.cpp | 2k | 2,816.0 | 115.5 | — | github.com | 2025-08-28 |
| Apple M4 Max (32-core GPU) | MXFP4 | llama.cpp | 2k | 1,277.0 | 92.4 | — | github.com | 2025-08-18 |
Frequently asked questions
How much VRAM does gpt-oss-20b need?
At MXFP4 the file is 12.11 GB; with 8k context a graphics card needs about 12.3 GB (weights 11.48 GB, KV cache 0.2 GB, compute buffer 0.6 GB). Part of the file stays off the graphics card: llama.cpp keeps the input embedding table in system RAM. A Mac counts the whole file, so it needs a little more.
What is the cheapest GPU that runs gpt-oss-20b well?
Cheapest that runs it great: Radeon RX 9060 XT 16GB at $449 (street price as of 2026-08-09), est. 78.0 tok/s (62.4–93.6, calibrated estimate ±20%). Cheaper and still Runs well: GeForce RTX 3060 12GB at $250 (street price as of 2026-08-09), est. 63.9 tok/s (44.8–83.1, theoretical estimate ±30%).
Can gpt-oss-20b run on an 8 GB, 16 GB or 24 GB GPU?
At MXFP4 with 8k context: GeForce RTX 4060 — Runs well, est. 36.9 tok/s (25.9–48.0, theoretical estimate ±30%); GeForce RTX 5060 Ti 16GB — Runs great, 111.6 tok/s measured at 2k context (1 run; 8k estimate 104.3 tok/s, 83.5–125.2, calibrated estimate ±20%); GeForce RTX 3090 — Runs great, 162.0 tok/s measured at 2k context (1 run; 8k estimate 201.1 tok/s, 160.9–241.3, calibrated estimate ±20%).
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.