# T4-Bench > Measured benchmarks of image, video, language, agent and fine-tuning workloads on a Kaggle 2x Tesla T4 notebook. The figures come from our own runs. This site helps people and AI agents answer one question: what runs well on two Kaggle T4 GPUs? ## Key facts - Hardware: 2x NVIDIA Tesla T4 15 GB, 31 GB RAM, 4 vCPU, CUDA 13, sm_75 Turing (no BF16 or FP8 hardware). - Fastest image model tested: FLUX.2 klein 4B, 374 images/hour at 1024x1024. - Fastest video model tested: LTX-Video 2B distilled, 27.6 seconds per warm 49-frame clip. - Fastest chat model tested: LFM2.5-8B-A1B, 119 tokens/second, one T4. - Best small agent tested: Spark 4B, 3 of 3 agent tasks passed. - Leanest fine-tuning method tested: Unsloth NF4, 7.1 GB versus 14.9 GB for stock transformers. - Source notes: https://blog.sdad.pro/ ## Read the full data - [Complete benchmark tables](https://t4b.sdad.pro/benchmarks.html): all models, metrics and units in plain HTML. - [Full text for agents](https://t4b.sdad.pro/llms-full.txt): every table as plain text. - [Interactive homepage](https://t4b.sdad.pro/): visual charts and diagrams. ## Repository - https://github.com/instax-dutta/T4-Bench ## How to cite T4-Bench, "Kaggle 2x Tesla T4 benchmark results", 2026. https://t4b.sdad.pro/