{"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","schema":"tlb.runs/v1","runs":[{"model":"gemma-3-27b","hardware":"rtx-5090-32g","quant":"Q4_K_M","runtime":"llama.cpp (CUDA)","test":"tg128 @ 8k depth","tok_s":58.2,"min":56.9,"max":59,"repeats":5,"flag":"sample","date":"2026-10-12","note":"Placeholder figure from the approved mockup; not a measurement."},{"model":"gemma-3-27b","hardware":"rtx-4090-24g","quant":"Q4_K_M","runtime":"llama.cpp (CUDA)","test":"tg128 @ 8k depth","tok_s":33.4,"min":32.8,"max":33.9,"repeats":5,"flag":"sample","date":"2026-10-12","note":"Placeholder figure from the approved mockup; not a measurement."},{"model":"gemma-3-27b","hardware":"rtx-3090-24g-used","quant":"Q4_K_M","runtime":"llama.cpp (CUDA)","test":"tg128 @ 8k depth","tok_s":30.2,"min":29.8,"max":30.6,"repeats":5,"flag":"sample","date":"2026-10-12","note":"Placeholder figure from the approved mockup; not a measurement."},{"model":"gemma-3-27b","hardware":"rx-7900-xtx-24g","quant":"Q4_K_M","runtime":"llama.cpp (Vulkan)","test":"tg128 @ 8k depth","tok_s":27.1,"min":26.2,"max":27.6,"repeats":5,"flag":"sample","date":"2026-10-12","note":"Placeholder figure from the approved mockup; not a measurement."},{"model":"gemma-3-27b","hardware":"macbook-pro-m4-max-128g","quant":"Q4_K_M","runtime":"llama.cpp (Metal)","test":"tg128 @ 8k depth","tok_s":13.9,"min":13.5,"max":14.2,"repeats":5,"flag":"sample","date":"2026-10-12","note":"Placeholder figure from the approved mockup; not a measurement."},{"model":"gemma-3-27b","hardware":"dgx-spark-128g","quant":"Q4_K_M","runtime":"llama.cpp (CUDA)","test":"tg128 @ 8k depth","tok_s":7.4,"min":7.2,"max":7.6,"repeats":5,"flag":"sample","date":"2026-10-12","note":"Placeholder figure from the approved mockup; not a measurement."},{"model":"gemma-3-27b","hardware":"ryzen-ai-max-395-128g","quant":"Q4_K_M","runtime":"llama.cpp (Vulkan)","test":"tg128 @ 8k depth","tok_s":6.9,"min":6.6,"max":7.1,"repeats":5,"flag":"sample","date":"2026-10-12","note":"Placeholder figure from the approved mockup; not a measurement."},{"model":"gemma-3-27b","hardware":"rtx-3090-24g-used","quant":"Q4_K_M","runtime":"llama.cpp","test":"tg128 @ depth 0","depth":0,"tok_s":30.2,"flag":"sample","date":"2026-10-12","note":"Placeholder depth curve from the mockup."},{"model":"gemma-3-27b","hardware":"rtx-3090-24g-used","quant":"Q4_K_M","runtime":"llama.cpp","test":"tg128 @ depth 4096","depth":4096,"tok_s":29.1,"flag":"sample","date":"2026-10-12","note":"Placeholder depth curve from the mockup."},{"model":"gemma-3-27b","hardware":"rtx-3090-24g-used","quant":"Q4_K_M","runtime":"llama.cpp","test":"tg128 @ depth 16384","depth":16384,"tok_s":26,"flag":"sample","date":"2026-10-12","note":"Placeholder depth curve from the mockup."},{"model":"gemma-3-27b","hardware":"rtx-3090-24g-used","quant":"Q4_K_M","runtime":"llama.cpp","test":"tg128 @ depth 32768","depth":32768,"tok_s":23.1,"flag":"sample","date":"2026-10-12","note":"Placeholder depth curve from the mockup."},{"model":"gemma-3-27b","hardware":"macbook-pro-m4-max-128g","quant":"Q4_K_M","runtime":"llama.cpp","test":"tg128 @ depth 0","depth":0,"tok_s":13.9,"flag":"sample","date":"2026-10-12","note":"Placeholder depth curve from the mockup."},{"model":"gemma-3-27b","hardware":"macbook-pro-m4-max-128g","quant":"Q4_K_M","runtime":"llama.cpp","test":"tg128 @ depth 4096","depth":4096,"tok_s":13.1,"flag":"sample","date":"2026-10-12","note":"Placeholder depth curve from the mockup."},{"model":"gemma-3-27b","hardware":"macbook-pro-m4-max-128g","quant":"Q4_K_M","runtime":"llama.cpp","test":"tg128 @ depth 16384","depth":16384,"tok_s":11.2,"flag":"sample","date":"2026-10-12","note":"Placeholder depth curve from the mockup."},{"model":"gemma-3-27b","hardware":"macbook-pro-m4-max-128g","quant":"Q4_K_M","runtime":"llama.cpp","test":"tg128 @ depth 32768","depth":32768,"tok_s":9.4,"flag":"sample","date":"2026-10-12","note":"Placeholder depth curve from the mockup."},{"model":"gemma-3-27b","hardware":"ryzen-ai-max-395-128g","quant":"Q4_K_M","runtime":"llama.cpp","test":"tg128 @ depth 0","depth":0,"tok_s":6.9,"flag":"sample","date":"2026-10-12","note":"Placeholder depth curve from the mockup."},{"model":"gemma-3-27b","hardware":"ryzen-ai-max-395-128g","quant":"Q4_K_M","runtime":"llama.cpp","test":"tg128 @ depth 4096","depth":4096,"tok_s":6.5,"flag":"sample","date":"2026-10-12","note":"Placeholder depth curve from the mockup."},{"model":"gemma-3-27b","hardware":"ryzen-ai-max-395-128g","quant":"Q4_K_M","runtime":"llama.cpp","test":"tg128 @ depth 16384","depth":16384,"tok_s":5.6,"flag":"sample","date":"2026-10-12","note":"Placeholder depth curve from the mockup."},{"model":"gemma-3-27b","hardware":"ryzen-ai-max-395-128g","quant":"Q4_K_M","runtime":"llama.cpp","test":"tg128 @ depth 32768","depth":32768,"tok_s":4.8,"flag":"sample","date":"2026-10-12","note":"Placeholder depth curve from the mockup."},{"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":"qwen3-30b-a3b","hardware":"rtx-3090-24g-used","quant":"Q4_K_M","runtime":"llama.cpp (CUDA)","test":"tg128 @ 8k depth","tok_s":112,"flag":"sample","date":"2026-10-12","note":"Placeholder figure from the approved mockup; not a measurement."},{"model":"qwen3-14b","hardware":"rtx-3090-24g-used","quant":"Q4_K_M","runtime":"llama.cpp (CUDA)","test":"tg128 @ 8k depth","tok_s":52.3,"flag":"sample","date":"2026-10-12","note":"Placeholder figure from the approved mockup; not a measurement."},{"model":"mistral-small-3.2-24b","hardware":"rtx-3090-24g-used","quant":"Q4_K_M","runtime":"llama.cpp (CUDA)","test":"tg128 @ 8k depth","tok_s":34.1,"flag":"sample","date":"2026-10-12","note":"Placeholder figure from the approved mockup; not a measurement."},{"model":"llama-2-7b","hardware":"macbook-pro-m4-max-128g","quant":"Q4_0","runtime":"llamafile (Metal)","test":"LocalScore generation","tok_s":62.5,"flag":"sourced","source":"https://www.localscore.ai/model/1248","date":"2026-10-12","note":"Third-party measurement used to calibrate η (brief §4.3): implies η ≈ 0.43."},{"model":"llama-2-7b","hardware":"rtx-5090-32g","quant":"Q4_0","runtime":"llama.cpp (CUDA)","test":"tg128","tok_s":290.02,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/discussions/15013","note":"Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table.","min":288.92,"max":291.12},{"model":"llama-2-7b","hardware":"rtx-4090-24g","quant":"Q4_0","runtime":"llama.cpp (CUDA)","test":"tg128","tok_s":186.21,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/discussions/15013","note":"Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table.","min":186.08,"max":186.34},{"model":"llama-2-7b","hardware":"rtx-3090-24g-used","quant":"Q4_0","runtime":"llama.cpp (CUDA)","test":"tg128","tok_s":158.16,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/discussions/15013","note":"Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table.","min":157.95,"max":158.37},{"model":"llama-2-7b","hardware":"rtx-5060-ti-16g","quant":"Q4_0","runtime":"llama.cpp (CUDA)","test":"tg128","tok_s":90.94,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/discussions/15013","note":"Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table.","min":90.92,"max":90.96},{"model":"llama-2-7b","hardware":"rtx-4060-ti-16g","quant":"Q4_0","runtime":"llama.cpp (CUDA)","test":"tg128","tok_s":63.86,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/discussions/15013","note":"Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table. Measured on the 8 GB variant (same 288 GB/s bus); treat as indicative for the 16 GB card.","min":63.85,"max":63.87},{"model":"llama-2-7b","hardware":"dgx-spark-128g","quant":"Q4_0","runtime":"llama.cpp (CUDA)","test":"tg128","tok_s":57.21,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/discussions/15013","note":"Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table.","min":57.15,"max":57.27},{"model":"llama-2-7b","hardware":"rx-7900-xtx-24g","quant":"Q4_0","runtime":"llama.cpp (Vulkan)","test":"tg128","tok_s":182.63,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/discussions/10879","note":"Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table.","min":182.41,"max":182.85},{"model":"llama-2-7b","hardware":"ryzen-ai-max-395-128g","quant":"Q4_0","runtime":"llama.cpp (Vulkan)","test":"tg128","tok_s":51.81,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/discussions/10879","note":"Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table. Device reported as \"AMD Ryzen AI Max+ 300 Series\" (Radeon 8060S).","min":51.68,"max":51.94},{"model":"llama-2-7b","hardware":"macbook-pro-m4-max-128g","quant":"Q4_0","runtime":"llama.cpp (Metal)","test":"tg128","tok_s":83.06,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/discussions/4167","note":"Llama 2 7B Q4_0, llama-bench tg128, M4 Max 40-core GPU (546 GB/s), pinned commit 8e672ef (Nov 2023) for comparability; newer builds are faster."},{"model":"llama-2-7b","hardware":"mac-studio-m5-max","quant":"Q4_0","runtime":"llama.cpp (Metal)","test":"tg128","tok_s":119.92,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/discussions/4167","note":"Llama 2 7B Q4_0, llama-bench tg128, measured on the M5 Max 40-core GPU (614 GB/s) option, commit c1d0e7a. The base 32-core (460 GB/s) model will be slower."},{"model":"llama-2-7b","hardware":"mac-studio-m5-ultra","quant":"Q4_0","runtime":"llama.cpp (Metal)","test":"tg128","tok_s":179.1,"flag":"sourced","date":"2026-10-12","source":"https://github.com/ggml-org/llama.cpp/discussions/4167","note":"Llama 2 7B Q4_0, llama-bench tg128, measured on the M5 Ultra 80-core GPU option (1.2 TB/s), commit c1d0e7a. The base 64-core model may be slightly slower."},{"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},{"model":"gpt-oss-120b","hardware":"dgx-spark-128g","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg32 @ depth 0","tok_s":58.72,"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":58.52,"max":58.92,"depth":0},{"model":"gpt-oss-120b","hardware":"dgx-spark-128g","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg32 @ depth 4096","tok_s":55.67,"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":55.32,"max":56.02,"depth":4096},{"model":"gpt-oss-120b","hardware":"dgx-spark-128g","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg32 @ depth 8192","tok_s":52.87,"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":52.44,"max":53.3,"depth":8192},{"model":"gpt-oss-120b","hardware":"dgx-spark-128g","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg32 @ depth 16384","tok_s":49.45,"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":49.25,"max":49.65,"depth":16384},{"model":"gpt-oss-120b","hardware":"dgx-spark-128g","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg32 @ depth 32768","tok_s":42.76,"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":42.62,"max":42.9,"depth":32768},{"model":"gpt-oss-120b","hardware":"dgx-spark-128g","quant":"MXFP4","runtime":"llama.cpp (CUDA)","test":"tg32 @ 8k depth","tok_s":52.87,"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":52.44,"max":53.3},{"model":"qwen3-coder-30b-a3b-instruct","hardware":"dgx-spark-128g","quant":"Q8_0","runtime":"llama.cpp (CUDA)","test":"tg32 @ 8k depth","tok_s":48.02,"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. Model is Qwen3-Coder-30B-A3B-Instruct. Catalogue correction 2026-10-13: model identifier corrected from Qwen3-30B-A3B; shared architecture does not establish identical performance.","min":47.62,"max":48.42}]}