The Strontium blog.

News, methods and measurements from the team that trains, serves and runs the Strontium stack. Every claim comes with the evidence behind it.

No posts are published yet. The first ones are on their way.

What we write about.

Six areas account for most of our work. Most posts start from one of them.

Blackwell-native precision
Using the low-precision formats new GPUs run natively, where a model can take them without losing quality.
Diffusion serving systems
Batching, scheduling and caching for models that run many steps per output, so cost and latency stay predictable under real traffic.
Quality-gated routing
Checking every output against its brief, and sending work to the model and settings most likely to pass.
Reinforcement learning for media
Training models against the checks that decide whether an output is accepted.
Identity and personalization
Keeping a person, a product or a character the same across shots, cuts and formats.
Measurement science
What to measure, how to measure it, and how much evidence a claim needs.

Tell us what you work on and what you would want to measure first.