The lineup
Six sizes, from a ThinkPad to a home datacenter. Tell us what you're running and we'll show what fits.
- Parameters
- Design
- Size at Q4_K_M
- Runs on
- Status
File sizes are estimates at Q4_K_M, about 0.6 GB per billion parameters. "Fits" leaves ~20% headroom for context.
How we build
We'd rather you run a model than admire its parameter count.
Speed first
Every model gets its own multi-token prediction head, trained in from the start. Mid-size models use mixture-of-experts so only about a billion parameters work on each token.
Quantization is the target
Every tier is designed to be run at Q4_K_M. That's the version we test, tune, and size against real GPU memory, not round numbers.
Open, on a delay
The newest flagship lives behind our API. Once two newer generations ship, its weights go public, so older models are always open.
The water cycle
Every name comes from one system, so you can tell what something is before you open it.
Language models
Named for the sky, from frost to storm. Bigger model, bigger weather.
Rime · Wisp · Cirrus · Stratus · Nimbus · Maelstrom
Datasets
Named for moving water: what the models learn from.
Undertow
Encoders
Named for light in the atmosphere: models that see and hear.
Halo · Aurora · Peal · Corona