GET /v1/models

Lists every model the calling key may use: chat, embeddings, re-ranking, speech and decisions. The response follows the OpenAI schema and adds fields that help you pick a suitable model.

The list is per organisation. Owners and admins decide under Settings → Models which chat models are usable, so two keys from different organisations can see different lists. A catalogue model your organisation has hidden answers POST /v1/chat/completions with 404 model_not_found, and so does an unknown ID (see headers). Reload the list on a 404 instead of hard-wiring model IDs.

curl https://sovrgpt.com/api/v1/models \
  -H "Authorization: Bearer $SOVR_KEY"

Response

{
  "object": "list",
  "data": [
    {
      "id": "gemma-4-12b",
      "object": "model",
      "created": 1715520000,
      "owned_by": "sovrgpt",

      "kind": "chat",
      "tier": "default",
      "display_name": "Gemma 4 12B",
      "subtitle": "Alltag – mit sehr großem Kontext",
      "capabilities": {
        "reasoning": 3,
        "coding": 3,
        "vision": 4,
        "speed": 4,
        "german": 4
      },
      "accepts_vision": true,
      "accepts_tools": true,
      "effort_levels": [],
      "context_window": 262144,
      "license": "Apache 2.0",
      "attribution": "Powered by Gemma",
      "quality_index": 22,
      "quality_index_source": "Artificial Analysis Intelligence Index",
      "quality_index_as_of": "2026-08-08",
      "origin_vendor": "Google DeepMind",
      "origin_country": "US",
      "lifecycle": "active",
      "cold_start_hint": "warm <10 s, Cold-Start gemessen ~5,5–9,5 min (332–574 s)"
    },
    {
      "id": "stackit-qwen3.8-27b",
      "object": "model",
      "owned_by": "sovrgpt",
      "tier": "hosted",
      "display_name": "Qwen 3.8 27B",
      "subtitle": "Ohne Kaltstart · Betrieb in Deutschland",
      "license": "Apache 2.0"
    }
    /* … further models … */
  ]
}

Fields

Standard OpenAI fields

SovrGPT extensions

The text fields (subtitle, description, cold_start_hint) are returned in German. OpenAI clients ignore unknown fields; every extension is additive.

About the context window

This value is the model's hard ceiling for input plus output. You need it to size max_tokens sensibly.

Exceeding it does not announce itself. /v1/chat/completions forwards your request to the model server unchanged; if it is too long, it is silently truncated there. You get HTTP 200 and an answer that never saw the part of your prompt that was cut off. There is no error code for it. So do the arithmetic yourself:

input tokens  +  max_tokens   <   context_window

The spread is wide: from 16,384 (qwen3.5-9b-deepseek-v4, llama-3.1-8b) to 1,000,000 (tensorx-deepseek-v4.1-flash). Switching an integration to another model also switches its context budget, see Migration. Models with a lot of context include gemma-4-12b (262,144), stackit-qwen3.8-27b (262,144) and tensorx-deepseek-v4.1-flash (1,000,000).

About the quality index

The source is the Artificial Analysis Intelligence Index. It measures every model in the same test setup, so the values are comparable with each other, unlike figures from vendors' own model cards.

Two limits: the index is based on English-language tests and says nothing about German language quality. And it is a snapshot, which is why we always ship quality_index_as_of. Models it does not list have quality_index: null.

About the lifecycle (lifecycle)

When a model is withdrawn, it disappears from the list and /v1/chat/completions answers 404 model_withdrawn with the successor. If you plan for the long term, filter on lifecycle instead of hard-wiring model IDs.

Embedding, audio and decision models

These models are in the list too, with the matching kind: bge-m3, nomic-embed-code, stackit-qwen3-vl-embedding-8b and stackit-e5-mistral-7b (embedding), bge-reranker-base (rerank), sovr-decision-v1 and sovr-decision-v2 (decision), and the speech models.

The embedding models compared (POST /v1/embeddings):

Without model, and for the OpenAI names (text-embedding-3-small etc.), bge-m3 keeps doing the work, so your existing vectors stay in their space. Any other dimensions value than those in the table ends in a 400 (param: "dimensions") for the STACKIT embedders. They accept at most 2,048 texts per request. With stackit-e5-mistral-7b, prefix search queries with a task line (Instruct: … Query: …) and leave documents as they are. The STACKIT embedders only appear when they are set up on the server; otherwise /v1/embeddings answers 503 model_not_available for them.

The speech models:

{ "id": "supertonic-3",  "kind": "tts", "display_name": "Supertonic 3 — Text-to-Speech" }
{ "id": "cosyvoice-3",   "kind": "tts", "display_name": "CosyVoice 3 — Ausdrucksstarkes TTS + Stimmklon" }
{ "id": "voxtral-mini-tts", "kind": "tts", "display_name": "Voxtral Mini TTS — Stimmenkatalog" }
{ "id": "voxtral-mini-transcribe", "kind": "stt", "display_name": "Voxtral Mini Transcribe — Speech-to-Text" }

Pass the id as model to POST /v1/audio/speech or /v1/audio/transcriptions; see the Audio API. These entries do not depend on the organisation's model selection.

Your organisation's own models

If your organisation registers its own OpenAI-compatible endpoint under Settings → Provider keys (your own address, model name and display name), it appears here like this:

{
  "id": "org-model:8b2f…",
  "owned_by": "org",
  "kind": "chat",
  "tier": "remote",
  "is_remote": true,
  "accepts_tools": false
}

• Use the id with the org-model: prefix in /v1/chat/completions like a catalogue ID.
• owned_by: "org": the endpoint is yours.
• Requests to such a model go straight to the registered address, outside our operation. The privacy policy of that provider applies.
• accepts_tools is false; effort_levels and context_window are absent. We make no promises about third-party endpoints we have not verified.

Only enabled entries appear. A paused or deleted entry disappears from the list, and /v1/chat/completions answers 404 model_not_found for its ID.

Model not in the list?

Possible reasons, to be checked in this order:

1. The organisation has hidden it. Owners and admins choose under Settings → Models which models the team sees. This is the most common reason, and you cannot see it from the outside.
2. It is planned (lifecycle: "planned") but not yet provisioned. Such models never appear in the list.
3. It has been withdrawn. /v1/chat/completions then answers 404 model_withdrawn and names the successor.
4. One of the organisation's own models is paused or deleted, or the server lacks the encryption key. In that case, own models are not served at all.

Filters and pagination

There are no filters and no pagination; the list currently has about 30 entries. If pagination arrives, it will follow the OpenAI pattern (?limit=/?after=).

Caching

The list changes rarely. Cache it on the client for about an hour. We do not push changes; reload the list after a 404 (model_not_found, model_withdrawn).