model_key stringclasses 25
values | repo_id stringclasses 25
values | family stringclasses 17
values | tier stringclasses 1
value | params_b float64 0.36 7.62 | checkpoint_gib float64 1.04 7.49 | n_shards int64 1 3 | bytes_per_param float64 0.73 4.28 | experiment stringclasses 3
values | condition stringclasses 6
values | p50_ms float64 678 66.6k | p95_ms float64 690 66.7k | stdev_ms float64 7.5 8.74k | rsd float64 0 2.15 | n_trials int64 3 5 | throughput_mib_s float64 109 2.15k | gpu stringclasses 1
value | device_class stringclasses 1
value | reliable bool 2
classes |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
bloomz-560m | bigscience/bloomz-560m | bloom | small | 0.56 | 1.042 | 1 | 1.998 | checkpoint_format | pytorch_bin | 902.46 | 1,190.88 | 139.69 | 0.1548 | 5 | 1,182.34 | Tesla T4 | t4 | true |
qwen3-0.6b | Qwen/Qwen3-0.6B | qwen3 | small | 0.6 | 1.4 | 1 | 2.505 | checkpoint_format | pytorch_bin | 1,850.08 | 2,485.44 | 422.23 | 0.2282 | 5 | 774.89 | Tesla T4 | t4 | true |
gpt2-medium | openai-community/gpt2-medium | gpt2 | small | 0.355 | 1.416 | 1 | 4.283 | checkpoint_format | pytorch_bin | 1,166.8 | 1,244.53 | 39.12 | 0.0335 | 5 | 1,242.71 | Tesla T4 | t4 | true |
tinyllama-1.1b-v1.0 | TinyLlama/TinyLlama-1.1B-Chat-v1.0 | tinyllama | small | 1.1 | 2.049 | 1 | 2 | checkpoint_format | pytorch_bin | 1,988.33 | 2,160.62 | 231.27 | 0.1163 | 5 | 1,055.24 | Tesla T4 | t4 | true |
deepseek-coder-1.3b | deepseek-ai/deepseek-coder-1.3b-instruct | deepseek | small | 1.35 | 2.508 | 1 | 1.995 | checkpoint_format | pytorch_bin | 2,192.96 | 2,620.15 | 215.68 | 0.0984 | 5 | 1,171.11 | Tesla T4 | t4 | true |
pythia-1.4b | EleutherAI/pythia-1.4b | pythia | small | 1.4 | 2.729 | 1 | 2.093 | checkpoint_format | pytorch_bin | 3,503.16 | 4,721.1 | 598.47 | 0.1708 | 5 | 797.71 | Tesla T4 | t4 | true |
qwen2.5-1.5b | Qwen/Qwen2.5-1.5B-Instruct | qwen2.5 | small | 1.54 | 2.875 | 1 | 2.005 | checkpoint_format | pytorch_bin | 2,783.45 | 4,950.61 | 1,080.88 | 0.3883 | 5 | 1,057.68 | Tesla T4 | t4 | false |
qwen2.5-coder-1.5b | Qwen/Qwen2.5-Coder-1.5B-Instruct | qwen2.5-coder | small | 1.54 | 2.875 | 1 | 2.005 | checkpoint_format | pytorch_bin | 2,740.7 | 3,540.9 | 428 | 0.1562 | 5 | 1,074.18 | Tesla T4 | t4 | true |
gpt2-large | openai-community/gpt2-large | gpt2 | small | 0.774 | 3.024 | 1 | 4.195 | checkpoint_format | pytorch_bin | 3,598.93 | 5,390.23 | 835.21 | 0.2321 | 5 | 860.42 | Tesla T4 | t4 | true |
falcon3-1b | tiiuae/Falcon3-1B-Instruct | falcon3 | small | 1 | 3.11 | 1 | 3.339 | checkpoint_format | pytorch_bin | 3,594.21 | 6,537.42 | 1,394.48 | 0.388 | 5 | 886.05 | Tesla T4 | t4 | false |
smollm2-1.7b | HuggingFaceTB/SmolLM2-1.7B-Instruct | smollm2 | small | 1.71 | 3.188 | 1 | 2.002 | checkpoint_format | pytorch_bin | 3,392.74 | 5,465 | 980.37 | 0.289 | 5 | 962.21 | Tesla T4 | t4 | true |
bloom-1b7 | bigscience/bloom-1b7 | bloom | small | 1.7 | 3.208 | 1 | 2.026 | checkpoint_format | pytorch_bin | 3,475.35 | 7,047.18 | 1,786.19 | 0.514 | 5 | 945.23 | Tesla T4 | t4 | false |
deepseek-r1-distill-qwen-1.5b | deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | deepseek-r1 | small | 1.78 | 3.31 | 1 | 1.997 | checkpoint_format | pytorch_bin | 3,407.89 | 8,064.1 | 2,245.43 | 0.6589 | 5 | 994.59 | Tesla T4 | t4 | false |
qwen3-1.7b | Qwen/Qwen3-1.7B | qwen3 | small | 1.7 | 3.784 | 2 | 2.39 | checkpoint_format | pytorch_bin | 4,360.76 | 8,873.03 | 2,432.73 | 0.5579 | 5 | 888.56 | Tesla T4 | t4 | false |
granite-3.1-2b | ibm-granite/granite-3.1-2b-instruct | granite | small | 2.53 | 4.719 | 2 | 2.003 | checkpoint_format | pytorch_bin | 21,260.61 | 21,322.24 | 43.08 | 0.002 | 5 | 227.29 | Tesla T4 | t4 | true |
phi-2 | microsoft/phi-2 | phi | small | 2.78 | 5.178 | 2 | 2 | checkpoint_format | pytorch_bin | 23,573.88 | 23,775.34 | 148.26 | 0.0063 | 5 | 224.92 | Tesla T4 | t4 | true |
stablelm-zephyr-3b | stabilityai/stablelm-zephyr-3b | stablelm | small | 2.8 | 5.207 | 1 | 1.997 | checkpoint_format | pytorch_bin | 21,730.99 | 22,894.97 | 1,264.96 | 0.0582 | 5 | 245.36 | Tesla T4 | t4 | true |
pythia-2.8b | EleutherAI/pythia-2.8b | pythia | small | 2.8 | 5.294 | 1 | 2.03 | checkpoint_format | pytorch_bin | 23,810.82 | 23,824.29 | 11.02 | 0.0005 | 5 | 227.67 | Tesla T4 | t4 | true |
smollm3-3b | HuggingFaceTB/SmolLM3-3B | smollm3 | small | 3.08 | 5.728 | 2 | 1.997 | checkpoint_format | pytorch_bin | 25,872.54 | 25,885.11 | 17.18 | 0.0007 | 5 | 226.71 | Tesla T4 | t4 | true |
stablelm-2-1-6b | stabilityai/stablelm-2-1_6b-chat | stablelm | small | 1.64 | 6.126 | 2 | 4.011 | checkpoint_format | pytorch_bin | 27,647.71 | 27,660.89 | 49.69 | 0.0018 | 5 | 226.89 | Tesla T4 | t4 | true |
bloomz-560m | bigscience/bloomz-560m | bloom | small | 0.56 | 1.042 | 1 | 1.998 | checkpoint_format | safetensors | 695.19 | 918.93 | 105.77 | 0.1521 | 5 | 1,534.85 | Tesla T4 | t4 | true |
qwen3-0.6b | Qwen/Qwen3-0.6B | qwen3 | small | 0.6 | 1.4 | 1 | 2.505 | checkpoint_format | safetensors | 949.5 | 1,338.82 | 182.07 | 0.1918 | 5 | 1,509.84 | Tesla T4 | t4 | true |
gpt2-medium | openai-community/gpt2-medium | gpt2 | small | 0.355 | 1.416 | 1 | 4.283 | checkpoint_format | safetensors | 890.56 | 1,193.11 | 143.14 | 0.1607 | 5 | 1,628.17 | Tesla T4 | t4 | true |
tinyllama-1.1b-v1.0 | TinyLlama/TinyLlama-1.1B-Chat-v1.0 | tinyllama | small | 1.1 | 2.049 | 1 | 2 | checkpoint_format | safetensors | 1,315.04 | 1,755.38 | 214.39 | 0.163 | 5 | 1,595.52 | Tesla T4 | t4 | true |
deepseek-coder-1.3b | deepseek-ai/deepseek-coder-1.3b-instruct | deepseek | small | 1.35 | 2.508 | 1 | 1.995 | checkpoint_format | safetensors | 1,715.57 | 5,410.06 | 1,865.45 | 1.0874 | 5 | 1,496.99 | Tesla T4 | t4 | false |
pythia-1.4b | EleutherAI/pythia-1.4b | pythia | small | 1.4 | 2.729 | 1 | 2.093 | checkpoint_format | safetensors | 1,723.52 | 2,185.96 | 226.97 | 0.1317 | 5 | 1,621.39 | Tesla T4 | t4 | true |
qwen2.5-1.5b | Qwen/Qwen2.5-1.5B-Instruct | qwen2.5 | small | 1.54 | 2.875 | 1 | 2.005 | checkpoint_format | safetensors | 1,807.28 | 2,044.37 | 114.38 | 0.0633 | 5 | 1,628.97 | Tesla T4 | t4 | true |
qwen2.5-coder-1.5b | Qwen/Qwen2.5-Coder-1.5B-Instruct | qwen2.5-coder | small | 1.54 | 2.875 | 1 | 2.005 | checkpoint_format | safetensors | 2,423.93 | 3,570.22 | 725.26 | 0.2992 | 5 | 1,214.56 | Tesla T4 | t4 | true |
gpt2-large | openai-community/gpt2-large | gpt2 | small | 0.774 | 3.024 | 1 | 4.195 | checkpoint_format | safetensors | 2,128.77 | 4,392.09 | 1,099.29 | 0.5164 | 5 | 1,454.63 | Tesla T4 | t4 | false |
falcon3-1b | tiiuae/Falcon3-1B-Instruct | falcon3 | small | 1 | 3.11 | 1 | 3.339 | checkpoint_format | safetensors | 2,086.46 | 2,794.37 | 366.85 | 0.1758 | 5 | 1,526.34 | Tesla T4 | t4 | true |
smollm2-1.7b | HuggingFaceTB/SmolLM2-1.7B-Instruct | smollm2 | small | 1.71 | 3.188 | 1 | 2.002 | checkpoint_format | safetensors | 1,980.48 | 2,461.33 | 230.82 | 0.1165 | 5 | 1,648.35 | Tesla T4 | t4 | true |
bloom-1b7 | bigscience/bloom-1b7 | bloom | small | 1.7 | 3.208 | 1 | 2.026 | checkpoint_format | safetensors | 2,017.37 | 7,039.78 | 2,380.52 | 1.18 | 5 | 1,628.35 | Tesla T4 | t4 | false |
deepseek-r1-distill-qwen-1.5b | deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | deepseek-r1 | small | 1.78 | 3.31 | 1 | 1.997 | checkpoint_format | safetensors | 2,872.84 | 8,002.05 | 2,524.72 | 0.8788 | 5 | 1,179.82 | Tesla T4 | t4 | false |
qwen3-1.7b | Qwen/Qwen3-1.7B | qwen3 | small | 1.7 | 3.784 | 2 | 2.39 | checkpoint_format | safetensors | 2,394.41 | 3,889.72 | 682.38 | 0.285 | 5 | 1,618.28 | Tesla T4 | t4 | true |
granite-3.1-2b | ibm-granite/granite-3.1-2b-instruct | granite | small | 2.53 | 4.719 | 2 | 2.003 | checkpoint_format | safetensors | 4,069.39 | 6,108.13 | 1,083.3 | 0.2662 | 5 | 1,187.46 | Tesla T4 | t4 | true |
phi-2 | microsoft/phi-2 | phi | small | 2.78 | 5.178 | 2 | 2 | checkpoint_format | safetensors | 4,854.01 | 16,357.51 | 5,425.72 | 1.1178 | 5 | 1,092.35 | Tesla T4 | t4 | false |
qwen2.5-7b-awq | Qwen/Qwen2.5-7B-Instruct-AWQ | qwen2.5-awq | small | 7.62 | 5.188 | 2 | 0.731 | checkpoint_format | safetensors | 4,918.35 | 13,317.7 | 3,969.75 | 0.8071 | 5 | 1,080.14 | Tesla T4 | t4 | false |
qwen2.5-7b-gptq-int4 | Qwen/Qwen2.5-7B-Instruct-GPTQ-Int4 | qwen2.5-gptq | small | 7.62 | 5.192 | 2 | 0.732 | checkpoint_format | safetensors | 5,751.15 | 16,538.7 | 4,928.46 | 0.857 | 5 | 924.44 | Tesla T4 | t4 | false |
stablelm-zephyr-3b | stabilityai/stablelm-zephyr-3b | stablelm | small | 2.8 | 5.207 | 1 | 1.997 | checkpoint_format | safetensors | 20,583.48 | 21,414.84 | 830.9 | 0.0404 | 5 | 259.04 | Tesla T4 | t4 | true |
pythia-2.8b | EleutherAI/pythia-2.8b | pythia | small | 2.8 | 5.294 | 1 | 2.03 | checkpoint_format | safetensors | 6,273.69 | 9,939.74 | 2,155.88 | 0.3436 | 5 | 864.09 | Tesla T4 | t4 | false |
smollm3-3b | HuggingFaceTB/SmolLM3-3B | smollm3 | small | 3.08 | 5.728 | 2 | 1.997 | checkpoint_format | safetensors | 7,371.82 | 11,363.7 | 2,344.86 | 0.3181 | 5 | 795.66 | Tesla T4 | t4 | false |
stablelm-2-1-6b | stabilityai/stablelm-2-1_6b-chat | stablelm | small | 1.64 | 6.126 | 2 | 4.011 | checkpoint_format | safetensors | 8,701.54 | 14,337.83 | 3,662.89 | 0.4209 | 5 | 720.91 | Tesla T4 | t4 | false |
phi-3.5-mini | microsoft/Phi-3.5-mini-instruct | phi | small | 3.82 | 7.117 | 2 | 2 | checkpoint_format | safetensors | 24,946.31 | 27,088.06 | 1,088.52 | 0.0436 | 5 | 292.14 | Tesla T4 | t4 | true |
phi-3-mini-4k | microsoft/Phi-3-mini-4k-instruct | phi | small | 3.82 | 7.117 | 2 | 2 | checkpoint_format | safetensors | 17,158.02 | 27,727.25 | 8,742.37 | 0.5095 | 5 | 424.75 | Tesla T4 | t4 | false |
qwen3-4b | Qwen/Qwen3-4B | qwen3 | small | 4 | 7.492 | 3 | 2.011 | checkpoint_format | safetensors | 25,288.99 | 25,555.12 | 141.61 | 0.0056 | 5 | 303.37 | Tesla T4 | t4 | true |
bloomz-560m | bigscience/bloomz-560m | bloom | small | 0.56 | 1.042 | 1 | 1.998 | checkpoint_format | safetensors-nommap | 685.26 | 689.58 | 8.12 | 0.0118 | 5 | 1,557.08 | Tesla T4 | t4 | true |
qwen3-0.6b | Qwen/Qwen3-0.6B | qwen3 | small | 0.6 | 1.4 | 1 | 2.505 | checkpoint_format | safetensors-nommap | 925.59 | 1,116.07 | 94.92 | 0.1025 | 5 | 1,548.85 | Tesla T4 | t4 | true |
gpt2-medium | openai-community/gpt2-medium | gpt2 | small | 0.355 | 1.416 | 1 | 4.283 | checkpoint_format | safetensors-nommap | 912.33 | 1,110.16 | 102.23 | 0.1121 | 5 | 1,589.31 | Tesla T4 | t4 | true |
tinyllama-1.1b-v1.0 | TinyLlama/TinyLlama-1.1B-Chat-v1.0 | tinyllama | small | 1.1 | 2.049 | 1 | 2 | checkpoint_format | safetensors-nommap | 1,342.92 | 1,737.58 | 186.82 | 0.1391 | 5 | 1,562.39 | Tesla T4 | t4 | true |
deepseek-coder-1.3b | deepseek-ai/deepseek-coder-1.3b-instruct | deepseek | small | 1.35 | 2.508 | 1 | 1.995 | checkpoint_format | safetensors-nommap | 1,548.8 | 2,045.2 | 226.28 | 0.1461 | 5 | 1,658.18 | Tesla T4 | t4 | true |
pythia-1.4b | EleutherAI/pythia-1.4b | pythia | small | 1.4 | 2.729 | 1 | 2.093 | checkpoint_format | safetensors-nommap | 1,708.05 | 2,306.55 | 299.02 | 0.1751 | 5 | 1,636.08 | Tesla T4 | t4 | true |
qwen2.5-1.5b | Qwen/Qwen2.5-1.5B-Instruct | qwen2.5 | small | 1.54 | 2.875 | 1 | 2.005 | checkpoint_format | safetensors-nommap | 1,794.08 | 2,292.08 | 249.4 | 0.139 | 5 | 1,640.96 | Tesla T4 | t4 | true |
qwen2.5-coder-1.5b | Qwen/Qwen2.5-Coder-1.5B-Instruct | qwen2.5-coder | small | 1.54 | 2.875 | 1 | 2.005 | checkpoint_format | safetensors-nommap | 1,814.78 | 2,238.79 | 205.43 | 0.1132 | 5 | 1,622.23 | Tesla T4 | t4 | true |
gpt2-large | openai-community/gpt2-large | gpt2 | small | 0.774 | 3.024 | 1 | 4.195 | checkpoint_format | safetensors-nommap | 1,973.31 | 2,334.45 | 171.86 | 0.0871 | 5 | 1,569.23 | Tesla T4 | t4 | true |
falcon3-1b | tiiuae/Falcon3-1B-Instruct | falcon3 | small | 1 | 3.11 | 1 | 3.339 | checkpoint_format | safetensors-nommap | 1,966.19 | 2,539.21 | 282.26 | 0.1436 | 5 | 1,619.7 | Tesla T4 | t4 | true |
smollm2-1.7b | HuggingFaceTB/SmolLM2-1.7B-Instruct | smollm2 | small | 1.71 | 3.188 | 1 | 2.002 | checkpoint_format | safetensors-nommap | 1,991.74 | 2,593 | 279.76 | 0.1405 | 5 | 1,639.02 | Tesla T4 | t4 | true |
bloom-1b7 | bigscience/bloom-1b7 | bloom | small | 1.7 | 3.208 | 1 | 2.026 | checkpoint_format | safetensors-nommap | 1,951.27 | 2,333.12 | 177.08 | 0.0907 | 5 | 1,683.51 | Tesla T4 | t4 | true |
deepseek-r1-distill-qwen-1.5b | deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | deepseek-r1 | small | 1.78 | 3.31 | 1 | 1.997 | checkpoint_format | safetensors-nommap | 2,291.53 | 2,885.34 | 269.63 | 0.1177 | 5 | 1,479.12 | Tesla T4 | t4 | true |
qwen3-1.7b | Qwen/Qwen3-1.7B | qwen3 | small | 1.7 | 3.784 | 2 | 2.39 | checkpoint_format | safetensors-nommap | 2,330.11 | 2,908.15 | 290.51 | 0.1247 | 5 | 1,662.93 | Tesla T4 | t4 | true |
granite-3.1-2b | ibm-granite/granite-3.1-2b-instruct | granite | small | 2.53 | 4.719 | 2 | 2.003 | checkpoint_format | safetensors-nommap | 3,981.05 | 4,174.03 | 282.52 | 0.071 | 5 | 1,213.82 | Tesla T4 | t4 | true |
phi-2 | microsoft/phi-2 | phi | small | 2.78 | 5.178 | 2 | 2 | checkpoint_format | safetensors-nommap | 4,535.22 | 5,238.46 | 339.98 | 0.075 | 5 | 1,169.13 | Tesla T4 | t4 | true |
qwen2.5-7b-awq | Qwen/Qwen2.5-7B-Instruct-AWQ | qwen2.5-awq | small | 7.62 | 5.188 | 2 | 0.731 | checkpoint_format | safetensors-nommap | 5,614.19 | 7,816.06 | 1,369.67 | 0.244 | 5 | 946.27 | Tesla T4 | t4 | true |
qwen2.5-7b-gptq-int4 | Qwen/Qwen2.5-7B-Instruct-GPTQ-Int4 | qwen2.5-gptq | small | 7.62 | 5.192 | 2 | 0.732 | checkpoint_format | safetensors-nommap | 6,380.88 | 8,036.84 | 1,269.69 | 0.199 | 5 | 833.21 | Tesla T4 | t4 | true |
stablelm-zephyr-3b | stabilityai/stablelm-zephyr-3b | stablelm | small | 2.8 | 5.207 | 1 | 1.997 | checkpoint_format | safetensors-nommap | 6,279.41 | 13,709.66 | 3,620.08 | 0.5765 | 5 | 849.12 | Tesla T4 | t4 | false |
pythia-2.8b | EleutherAI/pythia-2.8b | pythia | small | 2.8 | 5.294 | 1 | 2.03 | checkpoint_format | safetensors-nommap | 4,392.32 | 5,097.48 | 397.84 | 0.0906 | 5 | 1,234.21 | Tesla T4 | t4 | true |
smollm3-3b | HuggingFaceTB/SmolLM3-3B | smollm3 | small | 3.08 | 5.728 | 2 | 1.997 | checkpoint_format | safetensors-nommap | 9,030.49 | 11,571.95 | 1,420.31 | 0.1573 | 5 | 649.52 | Tesla T4 | t4 | true |
stablelm-2-1-6b | stabilityai/stablelm-2-1_6b-chat | stablelm | small | 1.64 | 6.126 | 2 | 4.011 | checkpoint_format | safetensors-nommap | 7,954.9 | 10,001.57 | 1,801.97 | 0.2265 | 5 | 788.57 | Tesla T4 | t4 | true |
phi-3.5-mini | microsoft/Phi-3.5-mini-instruct | phi | small | 3.82 | 7.117 | 2 | 2 | checkpoint_format | safetensors-nommap | 11,431.3 | 20,865.54 | 5,211.51 | 0.4559 | 5 | 637.53 | Tesla T4 | t4 | false |
phi-3-mini-4k | microsoft/Phi-3-mini-4k-instruct | phi | small | 3.82 | 7.117 | 2 | 2 | checkpoint_format | safetensors-nommap | 8,225.27 | 11,290.91 | 2,247.82 | 0.2733 | 5 | 886.03 | Tesla T4 | t4 | true |
qwen3-4b | Qwen/Qwen3-4B | qwen3 | small | 4 | 7.492 | 3 | 2.011 | checkpoint_format | safetensors-nommap | 23,676.63 | 25,215.73 | 4,690.06 | 0.1981 | 5 | 324.02 | Tesla T4 | t4 | true |
bloomz-560m | bigscience/bloomz-560m | bloom | small | 0.56 | 1.042 | 1 | 1.998 | engine_init | transformers | 2,870.14 | 3,227.81 | 241.06 | 0.084 | 3 | 371.76 | Tesla T4 | t4 | true |
qwen3-0.6b | Qwen/Qwen3-0.6B | qwen3 | small | 0.6 | 1.4 | 1 | 2.505 | engine_init | transformers | 3,109.51 | 7,340.83 | 2,553.73 | 0.8213 | 3 | 461.04 | Tesla T4 | t4 | false |
gpt2-medium | openai-community/gpt2-medium | gpt2 | small | 0.355 | 1.416 | 1 | 4.283 | engine_init | transformers | 1,285.02 | 5,730.36 | 2,405.16 | 1.8717 | 3 | 1,128.37 | Tesla T4 | t4 | false |
tinyllama-1.1b-v1.0 | TinyLlama/TinyLlama-1.1B-Chat-v1.0 | tinyllama | small | 1.1 | 2.049 | 1 | 2 | engine_init | transformers | 3,825.68 | 6,858.9 | 2,452.1 | 0.641 | 3 | 548.45 | Tesla T4 | t4 | false |
deepseek-coder-1.3b | deepseek-ai/deepseek-coder-1.3b-instruct | deepseek | small | 1.35 | 2.508 | 1 | 1.995 | engine_init | transformers | 1,947.16 | 7,655.35 | 3,137.44 | 1.6113 | 3 | 1,318.95 | Tesla T4 | t4 | false |
pythia-1.4b | EleutherAI/pythia-1.4b | pythia | small | 1.4 | 2.729 | 1 | 2.093 | engine_init | transformers | 1,334.53 | 3,878.42 | 1,347.07 | 1.0094 | 3 | 2,094 | Tesla T4 | t4 | false |
qwen2.5-1.5b | Qwen/Qwen2.5-1.5B-Instruct | qwen2.5 | small | 1.54 | 2.875 | 1 | 2.005 | engine_init | transformers | 2,262.23 | 5,195.45 | 1,542.23 | 0.6817 | 3 | 1,301.37 | Tesla T4 | t4 | false |
qwen2.5-coder-1.5b | Qwen/Qwen2.5-Coder-1.5B-Instruct | qwen2.5-coder | small | 1.54 | 2.875 | 1 | 2.005 | engine_init | transformers | 4,861.99 | 10,770.98 | 3,882.37 | 0.7985 | 3 | 605.51 | Tesla T4 | t4 | false |
gpt2-large | openai-community/gpt2-large | gpt2 | small | 0.774 | 3.024 | 1 | 4.195 | engine_init | transformers | 4,213.05 | 10,072.76 | 3,759.56 | 0.8924 | 3 | 735 | Tesla T4 | t4 | false |
falcon3-1b | tiiuae/Falcon3-1B-Instruct | falcon3 | small | 1 | 3.11 | 1 | 3.339 | engine_init | transformers | 3,882.89 | 12,556.54 | 4,954.31 | 1.2759 | 3 | 820.17 | Tesla T4 | t4 | false |
smollm2-1.7b | HuggingFaceTB/SmolLM2-1.7B-Instruct | smollm2 | small | 1.71 | 3.188 | 1 | 2.002 | engine_init | transformers | 1,519.97 | 7,757.82 | 3,274.13 | 2.1541 | 3 | 2,147.74 | Tesla T4 | t4 | false |
bloom-1b7 | bigscience/bloom-1b7 | bloom | small | 1.7 | 3.208 | 1 | 2.026 | engine_init | transformers | 10,647.37 | 10,798.58 | 2,960.13 | 0.278 | 3 | 308.53 | Tesla T4 | t4 | true |
deepseek-r1-distill-qwen-1.5b | deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | deepseek-r1 | small | 1.78 | 3.31 | 1 | 1.997 | engine_init | transformers | 3,525.12 | 13,968.93 | 5,653.96 | 1.6039 | 3 | 961.51 | Tesla T4 | t4 | false |
qwen3-1.7b | Qwen/Qwen3-1.7B | qwen3 | small | 1.7 | 3.784 | 2 | 2.39 | engine_init | transformers | 8,692.28 | 18,968.11 | 7,204.68 | 0.8289 | 3 | 445.78 | Tesla T4 | t4 | false |
granite-3.1-2b | ibm-granite/granite-3.1-2b-instruct | granite | small | 2.53 | 4.719 | 2 | 2.003 | engine_init | transformers | 12,946.29 | 19,295.65 | 6,893.92 | 0.5325 | 3 | 373.25 | Tesla T4 | t4 | false |
phi-2 | microsoft/phi-2 | phi | small | 2.78 | 5.178 | 2 | 2 | engine_init | transformers | 17,282.65 | 19,939.73 | 6,935.56 | 0.4013 | 3 | 306.8 | Tesla T4 | t4 | false |
stablelm-zephyr-3b | stabilityai/stablelm-zephyr-3b | stablelm | small | 2.8 | 5.207 | 1 | 1.997 | engine_init | transformers | 19,660.7 | 20,809.66 | 3,447.14 | 0.1753 | 3 | 271.2 | Tesla T4 | t4 | true |
pythia-2.8b | EleutherAI/pythia-2.8b | pythia | small | 2.8 | 5.294 | 1 | 2.03 | engine_init | transformers | 19,414.53 | 21,810.34 | 3,112.51 | 0.1603 | 3 | 279.23 | Tesla T4 | t4 | true |
smollm3-3b | HuggingFaceTB/SmolLM3-3B | smollm3 | small | 3.08 | 5.728 | 2 | 1.997 | engine_init | transformers | 22,397.24 | 23,205.72 | 5,126.65 | 0.2289 | 3 | 261.88 | Tesla T4 | t4 | true |
stablelm-2-1-6b | stabilityai/stablelm-2-1_6b-chat | stablelm | small | 1.64 | 6.126 | 2 | 4.011 | engine_init | transformers | 19,685.11 | 21,764.45 | 1,300.91 | 0.0661 | 3 | 318.67 | Tesla T4 | t4 | true |
phi-3.5-mini | microsoft/Phi-3.5-mini-instruct | phi | small | 3.82 | 7.117 | 2 | 2 | engine_init | transformers | 25,219.89 | 30,470.21 | 3,045.41 | 0.1208 | 3 | 288.97 | Tesla T4 | t4 | true |
phi-3-mini-4k | microsoft/Phi-3-mini-4k-instruct | phi | small | 3.82 | 7.117 | 2 | 2 | engine_init | transformers | 15,268.86 | 22,811.14 | 6,656.83 | 0.436 | 3 | 477.3 | Tesla T4 | t4 | false |
qwen3-4b | Qwen/Qwen3-4B | qwen3 | small | 4 | 7.492 | 3 | 2.011 | engine_init | transformers | 32,412.47 | 32,641.33 | 250.79 | 0.0077 | 3 | 236.69 | Tesla T4 | t4 | true |
bloomz-560m | bigscience/bloomz-560m | bloom | small | 0.56 | 1.042 | 1 | 1.998 | storage_tier | local-nvme | 678.07 | 776.71 | 53.97 | 0.0796 | 5 | 1,573.59 | Tesla T4 | t4 | true |
qwen3-0.6b | Qwen/Qwen3-0.6B | qwen3 | small | 0.6 | 1.4 | 1 | 2.505 | storage_tier | local-nvme | 1,054.55 | 1,278.35 | 146.51 | 0.1389 | 5 | 1,359.44 | Tesla T4 | t4 | true |
gpt2-medium | openai-community/gpt2-medium | gpt2 | small | 0.355 | 1.416 | 1 | 4.283 | storage_tier | local-nvme | 922.55 | 928.6 | 7.5 | 0.0081 | 5 | 1,571.72 | Tesla T4 | t4 | true |
tinyllama-1.1b-v1.0 | TinyLlama/TinyLlama-1.1B-Chat-v1.0 | tinyllama | small | 1.1 | 2.049 | 1 | 2 | storage_tier | local-nvme | 1,367.22 | 1,845.32 | 241.16 | 0.1764 | 5 | 1,534.62 | Tesla T4 | t4 | true |
deepseek-coder-1.3b | deepseek-ai/deepseek-coder-1.3b-instruct | deepseek | small | 1.35 | 2.508 | 1 | 1.995 | storage_tier | local-nvme | 1,567.02 | 3,173.98 | 800.53 | 0.5109 | 5 | 1,638.91 | Tesla T4 | t4 | false |
pythia-1.4b | EleutherAI/pythia-1.4b | pythia | small | 1.4 | 2.729 | 1 | 2.093 | storage_tier | local-nvme | 1,744.79 | 3,175.54 | 711.31 | 0.4077 | 5 | 1,601.62 | Tesla T4 | t4 | false |
qwen2.5-1.5b | Qwen/Qwen2.5-1.5B-Instruct | qwen2.5 | small | 1.54 | 2.875 | 1 | 2.005 | storage_tier | local-nvme | 2,113.14 | 2,746.62 | 387.79 | 0.1835 | 5 | 1,393.19 | Tesla T4 | t4 | true |
LLM Container Cold-Start Benchmark
Measurements of how long it takes to bring a language model from cold storage to a state where it can serve its first token, across 25 open-weight models spanning 17 architecture families and 100.9 GiB of checkpoints, on a single NVIDIA T4.
Cold start is the latency a serverless or scale-to-zero inference platform pays when it has no warm replica. It decomposes into weight transfer from storage, deserialization into host memory, transfer to the accelerator, and engine bring-up. This dataset measures each phase separately rather than reporting one aggregate number, because the dominant term shifts with model size and setup -- and optimising the wrong term buys nothing.
Headline finding: cold start is superlinear in checkpoint size
Fitting load_ms = a x GiB^b gives b = 1.58 to 2.10 across conditions, where b = 1 would mean constant throughput. Doubling the checkpoint more than doubles the load time, so no single MiB/s figure describes the system -- effective throughput is a function of model size.
Concretely, throughput holds near 1535 MiB/s below 5.18 GiB and falls to 758 MiB/s above it -- a 2.02x collapse.
Throughput in the healthy regime: 1354 MiB/s (R2=0.868).
The most likely cause is host memory, not the storage device: the runtime had roughly 12-13 GiB of usable system RAM, and the memory-mapped load path depends on the OS page cache holding the checkpoint while tensors are materialised. This is stated as a hypothesis with an obvious test attached (re-run on a high-RAM instance and see whether the boundary moves), not as a settled conclusion.
Why it matters: a cold-start budget extrapolated from small models will be optimistic for large ones, and the gap widens with size. Below the boundary, storage bandwidth is the lever; above it, more storage bandwidth does not help, because the bottleneck has moved.
Other results
- safetensors vs legacy pickle: median 1.71x faster to load over 20 models with identical weights. The format costs no model quality and no hardware to change.
- Storage tier: a bandwidth-capped tier is roughly an order of magnitude slower than local NVMe on the same checkpoints.
- Engine bring-up is a comparable cost to weight transfer at these sizes, which is precisely the distinction that determines whether snapshot/restore techniques would pay off for a given deployment.
Files
| File | Contents |
|---|---|
observations.csv |
One row per (model, experiment, condition), joined to checkpoint size, parameter count, shard count and family. |
raw/runs.json |
The raw merged ledger, unmodified, so all statistics can be recomputed from source. |
cross_model_report.md |
The generated analysis: scaling fits, per-model tables, projections, noise profile, limitations. |
Columns
model_key, repo_id, family, tier, params_b, checkpoint_gib,
n_shards, bytes_per_param, experiment, condition, p50_ms, p95_ms,
stdev_ms, rsd, n_trials, throughput_mib_s, gpu, device_class,
reliable.
reliable is False where relative standard deviation is 30% or more. Those
rows are published rather than dropped -- a reader may reasonably pick a
different threshold -- but no conclusion above rests on them.
Method
- Each condition is repeated, warm-up runs are discarded, and p50/p95 are reported rather than a mean, because tail latency is what a scale-to-zero SLA is written against.
- The OS page cache is evicted between trials (
posix_fadvise(DONTNEED)on an unprivileged runtime), so reads are genuinely cold rather than served from RAM. - Runs were distributed across several Colab sessions coordinated through a shared Postgres ledger with atomic claims and epoch fencing, so no task was ever executed twice. Duplicate execution would not merely waste time -- two workers would contend for the same disk and corrupt the I/O measurement.
- Registry metadata (checkpoint sizes, shard counts, gating) was read from the Hugging Face API rather than estimated.
Limitations
- Free-tier hardware. A single consumer GPU with shared, contended host I/O. The ratios and the scaling behaviour transfer; the absolute milliseconds are not production numbers.
- Noise. Median RSD is around 15%, with a long tail concentrated in engine bring-up where JIT compilation varies run to run. Small effects here are not findings.
- Not measured: driver-level checkpoint/restore (proprietary GPU memory snapshotting), multi-GPU tensor-parallel loading, and real network-attached storage -- the remote tier is a software bandwidth cap, which bounds the penalty from below rather than reproducing it.
- Coverage gaps are listed at the top of the report.
Reproducing
pip install -e ".[distributed,plots]"
coldstart-fleet init --tier small --device-class t4
coldstart-fleet work --device cuda --device-class t4
coldstart-fleet merge --out ./merged
Citation
@misc{coldstart_bench_2026,
title = {LLM Container Cold-Start Benchmark},
year = {2026},
note = {Dataset: https://huggingface.co/datasets/ArchCoder/llm-cold-start-benchmark}
}
License
Apache-2.0. Measurements only; no model weights are redistributed.
- Downloads last month
- 32