vust

API guide · checked 2026-08-08

Yandex AI Studio: Models and Service Map

More than a model list: which API fits chat, embeddings, realtime speech, Wordstat, search, OCR and translation; which stable/RC/deprecated IDs the live catalog returned; and where schema validation and an explicit reasoning budget are mandatory.

25 resources · all transport smokes passed · no production routing changedIndependent guide; no live VUST integration claimed
Open VUST AI (not Yandex)
25 resources: 16 chat, 7 embeddings, 2 realtimeSchema/parsing and reasoning need explicit controlsWordstat, Search, OCR, Translate and SpeechKit are distinct surfaces

Choose the surface by job

OCR

Text in an image; make the user-data logging decision explicit.

Wordstat / Search API

Demand, associations, dynamics and regions — or separate web-result retrieval.

SpeechKit / Realtime

A dedicated speech service or two native realtime WebSocket models.

Chat / Embeddings

Generation and vectors; new embeddings are 256d v2, legacy aliases are 1536d.

Pareto adoption rules

  • Do not copy the whole catalog into runtime: require job-specific eval, cost and fallback first.
  • Do not mix 256d v2 vectors with legacy 1536d without an explicit index migration.
  • Do not put RC/deprecated aliases in new routes merely because a smoke passed.
  • Repeat model, price and schema checks immediately before release.

FAQ

What did Yandex AI Studio's live model catalog return?

On 2026-08-08 GET /v1/models returned 25 resources: 16 Chat Completions resources, seven embedding resources and two native Realtime WebSocket resources. Every resource passed its appropriate bounded transport smoke. That proves access and protocol compatibility, not production quality for every job.

Which Yandex model IDs are practical candidates?

The chat catalog included Alice AI LLM and Flash, DeepSeek v4 Flash, GPT-OSS 120B/20B, Qwen3 235B and Qwen3.6 35B, plus YandexGPT 5 Lite, Pro and 5.1 with stable, RC and deprecated aliases. Prefer stable IDs; inventory aliases are evidence, not an adoption recommendation.

What parsing and reasoning traps were found?

Yandex/Alice prompt-only JSON often arrived inside Markdown fences, so use schema enforcement or validated parsing. A simple Qwen3.6 smoke spent all 256 output tokens on hidden reasoning and returned no visible text; set reasoning_effort to none or an intentionally bounded value for simple jobs.

When should I use Wordstat, Search API, OCR, Translate or SpeechKit?

Use Wordstat for Russian query frequency, associations, dynamics and regional distribution; Search API for web-result retrieval; OCR for text in images; Translate for deterministic translation; SpeechKit for speech recognition and synthesis. Do not route these specialist jobs through a chat model merely because it is available.

Does VUST currently route user requests through Yandex AI Studio?

No. The 25-resource catalog and service smokes are candidate evidence. No VUST production route was changed, and this page does not claim that /chat or /models uses Yandex, Alice or YandexGPT.

Sources: official Yandex AI Studio documentation and VUST's own 2026-08-08 live smokes. Credentials and identifiers are never included in this page or logs. Official documentation.