Measured on Kaggle · 2026
What runs on
two Tesla T4s?
Measured numbers from image, video, language and training runs on the same free notebook. The figures come from our own runs.
Chat & coding models
How fast each model replies once it is loaded and warm.
Reply speed
Higher is better · tokens per secondWatch the ranking
Diagram from the same dataA token is roughly a word or part of a word.
Two agents at once
Same box, parallel requestsTool-use test
15 real tasks · score out of 100Agents & long memory
Small models that can use tools and remember long documents.
Agent tasks completed
Out of 3 · higher is betterTernary 27B reply speed
Tokens per second · higher is betterThe larger Spark model finishes more agent tasks, but needs more memory. Adding a second T4 nearly doubles reply speed for the ternary model.
Fine-tuning
Memory needed to train an 8B model with the same settings.
Memory used
GB · lower is betterWatch the memory race
Diagram from the same dataQuality was effectively identical between the two working methods, so memory use is the deciding factor.
Image generation
Four models, one engine, 1024×1024 pictures.
Pictures per hour
Higher is better · both GPUs busyVideo generation
Short clips. Wall time includes loading, generating and saving.
Seconds per clip
Lower is betterThe memory limit
Why the largest image model needs tricks to run at all.
Memory needed
GB · one T4 has 15 GBHow to read this
Every number in these tables comes from a saved run.
What we learned
Small and clever beats big on this machine: a 6B image model outruns a 33B one, and an 8B mixture-of-experts chat model outruns a 27B dense one.