Reference

Model aliases

Your app asks for a role. The gateway decides which model fills it, and can change that model without any change to your code.

Early accessBy invitationLast updated 5 October 2026

Why aliases

Hard-coded model names spread through code and go stale. With aliases, a model swap is one change at the gateway, tested once, instead of a release of every app. Each alias also carries its own timeouts and, for most aliases, its own fallback chain.

We do not publish which provider or model sits behind an alias. That is deliberate: it changes as we test faster or cheaper options, and your app should not depend on it.

Expert aliases

General-purpose roles. Your key’s list decides which of them you may call.

Expert aliases (early access)
Alias	Use it for	Timeout (request / stream)
`askshiva-fast`	The quickest replies: short chat, classification, routing. Hidden reasoning is off by default; send `reasoning_effort` to override.	30 s / 60 s
`askshiva-smart`	The default for general work and tool use.	120 s / 300 s
`askshiva-reason`	Deep reasoning and long analysis. Public data only; for personal data use `askshiva-reason-pd`.	240 s / 600 s
`askshiva-tool`	Function calling and strict JSON schemas.	90 s / 300 s
`askshiva-vision`	Understanding images sent in the messages.	120 s / 300 s
`askshiva-indic`	Text in Indian languages, for example Hindi, for adult users. Test your own language before you rely on it.	120 s / 300 s
`askshiva-embed`	Text embeddings, 1,536 dimensions.	60 s
`askshiva-image`	Image generation.	180 s
`askshiva-stt`	Speech to text.	180 s
`askshiva-tts`	Text to speech.	120 s

Data-class aliases

These aliases exist so that the kind of data decides where a request may go. Full rules are on the data classes page.

Data-class aliases (early access)
Alias	Class	Use it for
`askshiva-flash-pd`	Personal	Fast chat and extraction on personal data
`askshiva-reason-pd`	Personal	Deep reasoning on personal data (can take minutes)
`askshiva-strict`	Personal	The highest-stakes writing and review
`askshiva-flash-public`	Public	Bulk work on public data only
`askshiva-kids`	Children	Children’s products; one approved provider, no fallback
`askshiva-kids-fast`	Children	The faster children’s alias; one approved provider, no fallback
`askshiva-embed-oai`	Any class	Embeddings; supports the `dimensions` parameter

Local aliases

Served by models that run next to the gateway on our rented server, so the request does not go to an outside model provider. They have no fallback.

Local aliases (early access)
Alias	Route	Notes
`askshiva-embed-local`	`/embeddings`	1,024 dimensions, many languages. Not interchangeable with `askshiva-embed` (1,536): never mix the two in one index.
`askshiva-rerank`	`/rerank`	Multilingual reranker; Cohere-style request body.
`askshiva-tts-local`	`/audio/speech`	Text to speech with its own voice names, which are not valid for `askshiva-tts`.

Planned Local large-language-model aliases for writing and vision are configured but switched off until our GPU server is back in service. They are not on any key today.

Existing model names

Some apps already send specific model ids. While they move to aliases, the gateway accepts a small set of those ids as pass-through names, with their own fallbacks. Whether your key may use one is shown by GET /models. New code should use aliases.

Tips

Leave room for hidden reasoning. Some aliases reason before they answer, and those thinking tokens count against max_tokens. A limit that is too small returns an empty answer that is still charged. The gateway therefore raises a smaller limit to a floor: 4,000 tokens on askshiva-smart, 3,000 on askshiva-strict, and 8,000 on askshiva-reason and askshiva-reason-pd. It never lowers a limit you set and leaves an unset limit unset. Even at these floors, a hard question can occasionally use up the whole budget.
Strict JSON on personal data: prefer askshiva-flash-pd, or tool calls on askshiva-reason-pd. In our tests, strict JSON schemas on the personal reasoning alias sometimes returned a wrong field value.
Slow is normal for deep reasoning. A long answer from askshiva-reason-pd can take over two minutes. Do not put it behind a short hard timeout.
Unsupported parameters are dropped, not rejected, when a fallback model does not support them. Do not rely on an optional parameter being honoured by every model in a chain.