Models
What Eterial serves today, and which model to reach for.
Three models are available. All of them are open-weight, and all of them speak the same API — switching is a change of one string.
| Model | Reach for it when |
|---|---|
glm-5.3-flash | Serious coding. Frontier-class, and the strongest model here. |
minimax-m2.7 | Conversation: support agents, assistants, anything user-facing. |
deepseek-v4-flash | Volume: simple work, a lot of it, where the bill is the constraint. |
Choosing between them
glm-5.3-flash is the default, and the heavyweight: frontier-class coding
ability at open-weight prices, which is the whole argument for running inference
this way. Point it at code — implement this function, fix this bug, convert this
file — and at the architectural work and the large unfamiliar codebases the other
two are not for.
minimax-m2.7 is the one for anything that talks to a person. It holds a
conversation, follows a system prompt and calls tools, which is most of what an
AI support system does.
deepseek-v4-flash is the cheap one, and cheap by a wide margin rather than a
sliver — see Pricing for what the three currently cost. Reach for
it when the job is simple and there is a great deal of it: classification,
extraction, routing, tagging, summarising a queue. It reasons and calls tools, so
it can carry an agent loop; it is also the smallest of the three, so it rewards
work that does not need the extra capacity rather than work it will do badly and
cheaply.
One difference the ids do not show: glm-5.3-flash is the only one of the three
that takes images and PDFs. The other two advertise reasoning and tools and
nothing more, so a request carrying an image_url or a file content part routes
to GLM or to nothing at all. GET /v1/models is where to check that before
sending one.
The live list
GET /v1/models is authoritative — it returns exactly what you can send to right
now, along with the capabilities each model advertises. See
Models endpoint.
Model ids are short and stable. What changes underneath is which backend serves one, which is why the same id keeps working when the network is busy — see Routing.