AI Index

GPT OSS 20B

Language OpenAI Verified Released Aug 5, 2025
  • In Text Out Text
  • Reasoning· low, medium, high
  • Tool use
  • Structured output
  • Open weights

Prices

1 offering · prices as listed
Cloudflare Workers AI Global · USD @cf/openai/gpt-oss-20b
Conditions/ 1M tokensInputOutput
Standard $0.20$0.30

Prices from Cloudflare Workers AI pricing (checked Oct 31, 2026)

Standard: the price a request with no conditions pays (the headline price elsewhere on the site).

Details

Knowledge cutoff
Jun 2024
Licence
apache-2.0
Family
gpt-oss
Model ID

Not stated on OpenAI’s pages: max input, reasoning can be off, max thinking budget, aliases.

Run locally

Weights: openai/gpt-oss-20b
  • Parameters 20.9B
  • Active 3.6B · 32 experts, 4 per token
  • Native precision MXFP4
  • Layers 24
  • KV cache per token 48 KB
  • Config positions 128K (config.json max_position_embeddings)
  • API context 128K (maker’s API)
My hardware GeForce RTX 4090 24 GB · 1008 GB/s · 64 GB DDR5 dual channel · 89.6 GB/s
On your hardware · GeForce RTX 4090
Runs well MXFP4 (native) (automatic) 8K context 196–294 tok/sOne GPU
WeightsKV cacheOverheadUsable
Needs
≈ 13.6 GB (weights 12.8 (maker’s file) + KV cache 0.2 + overhead 0.6)
Usable
21.6 GB (90% of its memory)
Generation speed
196–294 tok/s (reads the active 3.6B per token)
Longest context
128K (the model’s limit)
First token
about 0.12–0.17 s for a 2K prompt
PrecisionWeightsNeeds at 8KOn your hardwareLongest context
MXFP4 (native)maker’s file12.8 GB≈ 13.6Runs well 196–294 tok/s128K
Q3_K_Mquality drops9.5 GB≈ 10.2Runs well 257–385 tok/s128K
IQ3_XXSquality drops7.5 GB≈ 8.1Runs well 319–478 tok/s128K
Q2_Kquality drops7.3 GB≈ 8.0Runs well 324–486 tok/s128K
IQ2_XXSquality drops5.0 GB≈ 5.7Runs well 445–668 tok/s128K

Click a row to choose a precision. “Maker’s file” sizes are the maker’s own files; others are parameters × bits per weight.

Compared with other hardware

Each at the highest precision it runs. Click one for details.

Smallest that runs it (MXFP4 (native), 8K): GPU GeForce RTX 5080; unified memory Apple M5 Pro (16-core GPU) · 24 GB.

Run commands

vLLM
vllm serve openai/gpt-oss-20b --max-model-len 8192
Memory to fine-tune
Full≈ 314.1 GBDoesn’t fit your hardware
LoRA≈ 41.6 GBDoesn’t fit your hardware
QLoRA≈ 13.6 GBFits your hardware

Batch 1, 2,048-token sequences, gradient checkpointing; LoRA and QLoRA train rank-16 adapters on the attention projections.

Estimates; facts from the maker’s Hugging Face repositories, as of 2026-11-06. How it’s estimated

Cited scores

From third-party leaderboards, not measured by us
  • Epoch Capabilities Index 135

    #107 of 118 listed models · CI 132.4–137.5

  • LMArena Text 1,317

    #144 of 163 listed models · CI 1,311–1,324 · 10,416 votes

  • Epoch AI · GPQA Diamond 60.8%

    #110 of 168 listed models · medium effort · ±2.49

    Also listed: low effort 53.2% · high effort 46%

  • Epoch AI · OTIS Mock AIME 2024–2025 65.3%

    #98 of 159 listed models · medium effort · ±5.93

    Also listed: high effort 50.8% · low effort 40.3%

  • Epoch AI, 'Capabilities & benchmarking'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/benchmarks'. Licensed CC BY 4.0. · License · as of Nov 4 – 5, 2026
  • LMArena (Arena), leaderboard-dataset on Hugging Face, licensed CC BY 4.0 · License · as of Oct 28, 2026
  • EQ-Bench by Samuel J. Paech (eqbench.com; EQ-bench/EQ-bench-site on GitHub), MIT License. · License · as of Oct 28, 2026
  • changes we made

Changes

All changes →
  1. Now priced on Cloudflare Workers AI Input $0.20 · Output $0.30

Sources

OpenAI’s official policies: data use, privacy, terms →

OpenAI · United States · openai.com

Snapshots, the same family and other models from this provider