{"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":"gemma-3-27b","name":"Gemma 3 27B IT","org":"Google DeepMind","task":"general","hf_repo":"google/gemma-3-27b-it","gguf_repo":"bartowski/google_gemma-3-27b-it-GGUF","license":{"value":"Gemma Terms of Use","flag":"verified","source":"https://ai.google.dev/gemma/terms"},"datasheet":"TLB-M-0006","rev":"2026.10","as_of":"2026-10-12","summary":"The 24 GB-card workhorse. Five of every six layers only look at the last 1,024 tokens, so the KV cache stays small as context grows.","arch":{"params_total":{"value":27.43,"flag":"verified","unit":"B","source":"https://huggingface.co/api/models/unsloth/gemma-3-27b-it","note":"safetensors.total from the HF API. Read from an ungated mirror of the gated official repo; identical architecture."},"params_active":{"value":27.43,"flag":"verified","unit":"B","source":"https://huggingface.co/api/models/unsloth/gemma-3-27b-it","note":"Dense: all parameters active."},"layers":{"value":62,"flag":"verified","source":"https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json","note":"Read from an ungated mirror of the gated official repo; identical architecture."},"kv_heads":{"value":16,"flag":"verified","source":"https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json","note":"Read from an ungated mirror of the gated official repo; identical architecture."},"q_heads":{"value":32,"flag":"verified","source":"https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json","note":"Read from an ungated mirror of the gated official repo; identical architecture."},"head_dim":{"value":128,"flag":"verified","source":"https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json","note":"Read from an ungated mirror of the gated official repo; identical architecture."},"hidden_size":{"value":5376,"flag":"verified","source":"https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json","note":"Read from an ungated mirror of the gated official repo; identical architecture."},"context":{"value":131072,"flag":"verified","unit":"tokens","source":"https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json"},"sliding_window":{"value":1024,"flag":"verified","unit":"tokens","source":"https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json","note":"Local attention window."},"sliding_layers":{"value":52,"flag":"verified","source":"https://huggingface.co/unsloth/gemma-3-27b-it/blob/main/config.json","note":"Layers using sliding-window attention (from layer_types / sliding_window_pattern)."}},"quants":{"Q3_K_M":{"bpw":3.91,"file_gb":{"value":13.44,"flag":"verified","unit":"GB","source":"https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF/tree/main","note":"Sum of GGUF file sizes (Q3_K_M) from the HF API tree listing."},"kld":{"value":0.071,"flag":"sample","note":"Placeholder until KLD runs (brief §4.2)."}},"Q4_K_M":{"bpw":4.85,"file_gb":{"value":16.55,"flag":"verified","unit":"GB","source":"https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF/tree/main","note":"Sum of GGUF file sizes (Q4_K_M) from the HF API tree listing."},"kld":{"value":0.017,"flag":"sample","note":"Placeholder until KLD runs (brief §4.2)."}},"Q5_K_M":{"bpw":5.69,"file_gb":{"value":19.27,"flag":"verified","unit":"GB","source":"https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF/tree/main","note":"Sum of GGUF file sizes (Q5_K_M) from the HF API tree listing."},"kld":{"value":0.008,"flag":"sample","note":"Placeholder until KLD runs (brief §4.2)."}},"Q6_K":{"bpw":6.56,"file_gb":{"value":22.17,"flag":"verified","unit":"GB","source":"https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF/tree/main","note":"Sum of GGUF file sizes (Q6_K) from the HF API tree listing."},"kld":{"value":0.004,"flag":"sample","note":"Placeholder until KLD runs (brief §4.2)."}},"Q8_0":{"bpw":8.5,"file_gb":{"value":28.71,"flag":"verified","unit":"GB","source":"https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF/tree/main","note":"Sum of GGUF file sizes (Q8_0) from the HF API tree listing."},"kld":{"value":0.001,"flag":"sample","note":"Placeholder until KLD runs (brief §4.2)."}},"F16":{"bpw":16,"file_gb":{"value":54.03,"flag":"verified","unit":"GB","source":"https://huggingface.co/bartowski/google_gemma-3-27b-it-GGUF/tree/main","note":"Sum of GGUF file sizes (BF16) from the HF API tree listing."}}},"quality":{"arena_text":{"value":1357.7,"flag":"sourced","source":"https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset","note":"LMArena text, overall, entry 'gemma-3-27b-it': 1357.7 (95% CI 1354.1–1361.4, 46,672 votes), leaderboard published 2026-10-08. CC-BY-4.0."}}},"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."}]}