{"as_of":"2026-10-12","license":"CC-BY-4.0","license_url":"https://creativecommons.org/licenses/by/4.0/","cite_as":"The Local Bench (thelocalbench.com), data as of 2026-10-12, CC BY 4.0","source_url":"https://thelocalbench.com","model":{"id":"gpt-oss-20b","name":"gpt-oss-20b","org":"OpenAI","task":"general","hf_repo":"openai/gpt-oss-20b","gguf_repo":"ggml-org/gpt-oss-20b-GGUF","license":{"value":"Apache-2.0","flag":"verified","source":"https://www.apache.org/licenses/LICENSE-2.0"},"datasheet":"TLB-M-0004","rev":"2026.10","as_of":"2026-10-12","summary":"A mixture-of-experts model shipped natively in MXFP4. Only 3.6B parameters are read per token, so it is fast on almost anything with 16 GB.","arch":{"params_total":{"value":20.91,"flag":"verified","unit":"B","source":"https://huggingface.co/api/models/openai/gpt-oss-20b","note":"safetensors.total from the HF API."},"params_active":{"value":3.6,"flag":"sourced","unit":"B","source":"https://huggingface.co/openai/gpt-oss-20b","note":"Active parameters per token from the model card."},"layers":{"value":24,"flag":"verified","source":"https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"},"kv_heads":{"value":8,"flag":"verified","source":"https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"},"q_heads":{"value":64,"flag":"verified","source":"https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"},"head_dim":{"value":64,"flag":"verified","source":"https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"},"hidden_size":{"value":2880,"flag":"verified","source":"https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"},"context":{"value":131072,"flag":"verified","unit":"tokens","source":"https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"},"sliding_window":{"value":128,"flag":"verified","unit":"tokens","source":"https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json","note":"Local attention window."},"sliding_layers":{"value":12,"flag":"verified","source":"https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json","note":"Layers using sliding-window attention (from layer_types / sliding_window_pattern)."},"experts":{"value":"32 experts, 4 active per token","flag":"verified","source":"https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json"}},"quants":{"MXFP4":{"bpw":4.25,"file_gb":{"value":12.11,"flag":"verified","unit":"GB","source":"https://huggingface.co/ggml-org/gpt-oss-20b-GGUF/tree/main","note":"Sum of GGUF file sizes (MXFP4) from the HF API tree listing."}}},"quality":{"arena_text":{"value":1287.4,"flag":"sourced","source":"https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","note":"LMArena text, overall, entry 'gpt-oss-20b': 1287.4 (95% CI 1281.0–1293.7, 10,418 votes), leaderboard published 2026-10-08. CC-BY-4.0."}}},"runs":[{"model":"gpt-oss-20b","hardware":"rtx-3090-24g-used","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg128 @ 8k depth","tok_s":128.4,"flag":"sample","date":"2026-10-12","note":"Placeholder figure from the approved mockup; not a measurement."},{"model":"gpt-oss-20b","hardware":"dgx-spark-128g","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg32 @ depth 0","tok_s":83.43,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md","note":"llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.","min":82.84,"max":84.02,"depth":0},{"model":"gpt-oss-20b","hardware":"dgx-spark-128g","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg32 @ depth 4096","tok_s":79.22,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md","note":"llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.","min":78.62,"max":79.82,"depth":4096},{"model":"gpt-oss-20b","hardware":"dgx-spark-128g","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg32 @ depth 8192","tok_s":75.22,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md","note":"llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.","min":74.17,"max":76.27,"depth":8192},{"model":"gpt-oss-20b","hardware":"dgx-spark-128g","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg32 @ depth 16384","tok_s":70.36,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md","note":"llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.","min":69.99,"max":70.73,"depth":16384},{"model":"gpt-oss-20b","hardware":"dgx-spark-128g","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg32 @ depth 32768","tok_s":61.65,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md","note":"llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.","min":61.35,"max":61.95,"depth":32768},{"model":"gpt-oss-20b","hardware":"dgx-spark-128g","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg32 @ 8k depth","tok_s":75.22,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/blob/master/benches/dgx-spark/dgx-spark.md","note":"llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark.","min":74.17,"max":76.27}]}