The engines that answer typed, calibrated decisions
Several engines answer the same questions and return the same answer shapes, in one of two
ways. A diffusion read seeds a canvas and denoises it once. A generative read
writes the answer template under a grammar. Either way nothing is parsed, so an answer
cannot be off-schema.
Model
Read
Notes
google/diffusiongemma-26b-a4b-it
diffusion
The default engine, and the most peaked distributions.
google/gemma-4-26b-a4b-it
generative
The sparse sibling of the default: 4B active, so a read costs a fraction of its size.
qwen/qwen3.8-27b
generative
The largest generative engine on this route.
qwen/qwen3.5-0.8b
generative
The smallest engine on this route. Decisions are less peaked than the default’s.
Each also accepts its Hugging Face spelling, for example Qwen/Qwen3.8-27B.
Generative reads are experimental. The diffusion engine reads probabilities straight
off the canvas. A generative engine writes the answer under a grammar, so the shape is
guaranteed but the calibration is not: on the smaller engines, measured probabilities sit
much closer to uniform and a confident-looking answer can still be wrong. Treat
confidence from a generative read as a weak signal, validate on your own labelled data
before you threshold on it, and prefer google/diffusiongemma-26b-a4b-it wherever the
decision matters.
There is no catch-all alias. Name the model you want. An id the deployment does not
serve is a 404, so the TypeSafe SDK’s jev-latest default has to be overridden. See
TypeSafe SDK Compatibility.
Which engines a deployment serves varies, and a decision reports the id that answered it,
so read the live list rather than hard-coding an id:
GET /typesafe/v1/models.
