AI models

This page lists every SovrGPT AI model: language models, embeddings and re-rankers, speech recognition and speech output, and the Decision Engine. You use them in chat or through the OpenAI-compatible API. Storage and AI processing take place in EU data centres. The exceptions are the optional web search (provider based in the US) and models your organisation brings itself.

The language models run in two operating modes. For you, the main difference is the wait:

• Self-operated: models we run on rented GPUs in EU data centres (and additionally in Iceland or Norway, EEA). The servers shut down during pauses. That keeps the price low, but the first request after a pause waits a few minutes for a cold start.
• Partner-operated: models that do not run, or do not run economically, in self-operation are operated by STACKIT in Germany and by a European partner data centre in Ireland/Finland. There, zero retention of inputs and outputs is guaranteed by contract. There is no cold start.

The Decision Engine has its own endpoint (POST /v1/decisions). Every other model is reachable through the same API; each table shows the operating mode.

The most important models

These models sit in the first row of the model overview ("Models" in the sidebar) and cover everyday work. If you would rather not choose, pick Auto: SovrGPT classifies your first question (everyday, reasoning, programming, interactive app) and picks a suitable model for it.

Qwen 3.8 27B is our standard Qwen and runs at STACKIT in Germany, without a cold start. The same model is also available in the partner data centre in Ireland/Finland (tensorx-qwen3.8-27b) and self-operated (qwen3.8-27b); both are listed under All language models. Self-operated, you wait around five minutes for the first answer after a pause, longer when GPUs are scarce. New chats start with Auto. Auto picks models without a cold start, for everyday work first Qwen 3.8 27B at STACKIT, then the same model in the partner data centre, and uses a self-operated one only when none of them is reachable.

The superscript marks are explained under Notes on the tables.

All language models

In the model overview, everything beyond the first row sits under "More models". Through the API every model is addressable in the same way. GET /api/v1/models returns the current state, including the fields quality_index, origin_vendor, origin_country and lifecycle (API reference).

Partner-operated (no cold start)

Context 131k tokens ³ unless stated otherwise.

Notes on individual models:

• Qwen 3.8 27B (stackit-qwen3.8-27b) is strong at code and tool calls and understands images. In our test run the first token arrived after about 120 ms ᵐ; it accepted a prompt of 257,000 tokens ᵐ. You steer the thinking effort with reasoning_effort (low, medium, high).
• GPT-OSS 120B always thinks. At thorough the answer needs a generous token budget, otherwise it all goes into the chain of thought. The first token arrives after about 130 ms ᵐ.
• Qwen 3.8 Max takes around 2.4 TB for its weights alone, which is why it is not economical to self-operate. enable_thinking: false is rejected.
• GLM 5.3 and GLM 5.3 Flash always think, so the model picker shows no thinking toggle for them. Both pass our tool test on all three paths (freely chosen, forced, streaming) and answer in under a second.
• Qwen 3.8 Flash has 125bn parameters, of which only 6bn compute per token; hence the price. Thinking can be switched off, and the reasoning part is then really empty. It is a different model from Qwen 3.8 27B. If you need the verified tool and coding quality of the 27B, stay with it.
• DeepSeek V4 Flash is the release of 31 July 2026 (DeepSeek-V4-Flash-0731). The withdrawn coder-max tier ran the older release DeepSeek-V4-Flash.
• DeepSeek V4.1 Flash has no reasoning part. A thinking toggle would have nothing to switch, so there is none.
• Gemma 4 31B understands images and runs in Germany. Among the self-operated models only vision can read images, with a cold start. Over 140 languages, thinking toggle, first token after about 210 ms ᵐ.
• Kimi K2.7 Code is the alternative to coder-mini when the cold start gets in the way.

Self-operated (rented GPUs, with cold start)

Notes on individual models:

• gemma-4-12b is the first choice for interactive chats and German text. With 262k context it takes long documents in one piece, and it understands images.
• qwen3.8-27b ("Qwen 3.8 27B · eigener Betrieb", i.e. self-operated) is a dense 27B model for repository-wide code tasks and tool calling, with fluent German. It is the same model as our standard Qwen stackit-qwen3.8-27b, here on our own hardware and with a cold start. Its context also determines how much of an uploaded PDF fits into the conversation. The older ID qwen3.6-27b as well as qwen-balanced, qwen-premium and Qwen/Qwen3.8-27B-FP8 keep pointing to this model.
• qwen, qwen-default, Qwen/Qwen3.8-27B and the older Qwen names (such as Qwen/Qwen3.6-27B, qwen2.5-7b, qwen-3-30b) point to the standard Qwen stackit-qwen3.8-27b. On a server where STACKIT is not set up, they land on qwen3.8-27b. If you specifically want self-operation, name qwen3.8-27b.
• qwen3.5-9b-deepseek-v4 always thinks and suits mathematics, logic and step-by-step evaluations. The API returns the chain of thought in the reasoning_content field, not as <think>…</think> inside content. In chat it appears as an expandable "chain of thought".
• gemma-4-26b-a4b reads PNG and JPG, uses tools even together with an image and has a thinking toggle.
• qwen3.6-35b-a3b is the fast, low-cost coding model (MoE, 35bn parameters, 3bn of them active, FP8) with tool calling and adjustable thinking effort. See Coding in the IDE.
• llama-3.1-8b serves Llama-specific workloads, comparisons or requirements such as "Built with Llama". It has no tools (why).

Embeddings and re-rankers (RAG)

For retrieval-augmented generation (RAG) there are embedding and rerank models. Most of them are self-operated on rented GPU servers in EU data centres (and additionally in Iceland or Norway, EEA), two embedders run at STACKIT in Germany. You reach them through POST /api/v1/embeddings (OpenAI-compatible) and POST /api/v1/rerank (Cohere-compatible).

• bge-m3 is the default embedder: multilingual, with good retrieval quality in German. Without model, and for OpenAI names such as text-embedding-3-small, this model always does the work, so existing vectors stay in their space.
• nomic-embed-code specialises in code, for example for repository search.
• stackit-qwen3-vl-embedding-8b accepts up to 32,000 tokens per text and returns 4096 dimensions. With the dimensions parameter you get shorter vectors (256, 384, 768, 1024 or 1536), truncated and re-normalised. It is the same embedder the knowledge base works with.
• stackit-e5-mistral-7b always returns 4096 dimensions; dimensions only accepts 4096. According to its model card it is trained for texts of up to 4,096 tokens. Search queries get a task line in front (Instruct: … Query: …), documents do not.
• bge-reranker-base re-orders a candidate list by relevance to the query. It works well as a second stage after an embedding search.

Vectors from different embedders do not fit together. If you switch the model, regenerate all vectors. Details on the dimensions parameter: GET /v1/models.

The self-operated models shut down during pauses. Measured, bge-m3 then takes about a minute and the re-ranker about half a minute ᵐ; nomic-embed-code has not been measured. The two embedders at STACKIT have no cold start. In GET /api/v1/models they appear with kind: "embedding" or kind: "rerank".

An unknown model ID, including one of the wrong kind (a re-ranker sent to /v1/embeddings, a chat model sent to /v1/rerank), fails with 404 model_not_found before a GPU job starts. x-sovrgpt-model-served always names the model that did the work. The OpenAI names and HF paths listed above keep working.

Voice: speech-to-text and text-to-speech

In chat you dictate through the microphone icon, and the speaker button under every answer reads it aloud. Programmatically you use the OpenAI-compatible Audio API.

• Mistral (Voxtral Mini TTS) has been the default voice since 26 September 2026. The same engine speaks the designed German voices and the stock voices; if you pick a catalogue voice, SovrGPT uses this engine automatically. Output mp3, wav, opus or flac. Catalogue voices are selectable from the Starter plan or with the "Voices" org entitlement. Catalogue and samples: Voices.

• Supertonic 3 was the default voice until 26 September 2026. Today it speaks languages that neither Mistral nor Qwen speak, and on explicit request: an ONNX model with about 99M parameters, ten voices (M1–M5 male, F1–F5 female) and expression tags such as <breath> and <sigh> in the text. Output wav, flac or ogg, with no cold start. The service runs on a rented server in the Netherlands; the hosting provider is based in the US.

• CosyVoice 3 (Apache 2.0) speaks German natively and can do what Supertonic cannot:

  • Inline tags in the middle of the text: [laughter], [breath], <laughter>…</laughter>, <strong>…</strong> (emphasis).
  • Speaking style by instruction, for example emotion: "Sprich sehr traurig und langsam.".
  • Zero-shot voice cloning: a voice of your own from one reference clip (≤ 30 s), without training. Built-in voice: de-thorsten.

  You select it with provider: "cosyvoice". When warm it answers in about 6 s; after a pause a cold start of about 3 minutes ᵐ comes on top. Its output carries an AI label in the metadata and an inaudible watermark in the audio signal (AI labelling).

• Qwen3-TTS (Apache 2.0, provisional) speaks four designed voices of our own without the text going to Mistral. After a pause a cold start of just under a minute comes on top.

• Voxtral recognises speech in many languages and reliably in German. context_bias pulls recognition towards proper nouns and domain terms. format_text=1 returns a formatted transcript with paragraphs, numbered lists and a set-off salutation and closing; the raw transcript always comes along as text_raw.

• Whisper large-v3 transcribes 99 languages without the audio reaching a third-party speech service. Cold after a pause about 20 s, warm a few seconds.

You choose the provider per request in the provider field: mistral (default), supertonic, cosyvoice or qwen. Without a value, Mistral applies; if Mistral does not speak the language, Qwen takes over, otherwise Supertonic. In GET /api/v1/models all of them appear with kind: "tts" or kind: "stt". Examples with curl and the SDK: Audio API.

Decisions: Decision Engine

The Decision Engine produces no text. It takes a state (text, JSON, a conversation, optionally images) and questions with fixed answers (choice, yes/no or score), and returns probabilities over exactly those answers. The model runs one pass per question, or two in parallel when choice questions get the double check. That is why a decision comes back in milliseconds when warm.

The times are measurements without the double check (warm, four answer options), not a guarantee; we quote none for requests with images. Model versions, limits and prices may still change.

Both versions run on open base models; we decide which runtime computes them. Owners or admins choose your organisation's default under "Decision Engine". Through the API you use POST /v1/decisions with the decisions scope; GET /v1/models lists both with kind: "decision". In chat the tool is called "Entscheiden" (decide), via MCP sovrgpt_decide. Ten presets with examples are listed under Use cases.

Answers are not calibrated yet. Each carries a recommendation (recommendation: act or review) with warning signals (reasons); what follows from it is up to your workflow. The default is policy.mode: "threshold" with a double check of choice questions, which doubles their input tokens. The other modes and their cost are described in the API reference. We will quote accuracies only after calibration with a separate test set.

Withdrawn models

Through the API, SovrGPT answers an exact old ID with 404 model_withdrawn. The error message names the successor, and the x-sovrgpt-model-withdrawn header names the withdrawn ID. Chats that were pinned to a withdrawn model continue on the default model.

Role names point to the successors: coder, coding, qwen-coder and qwen3-coder to qwen3.6-35b-a3b, qwen-premium to qwen3.8-27b, and the old 30B names to stackit-qwen3.8-27b.

An unknown ID, for example one with a typo, fails with 404 model_not_found (Chat Completions).

Notes on the tables

ᵐ Measured at the live endpoint. The figure is the time you wait, so it includes waiting for a free GPU in the EU. When warm, every self-operated model usually starts answering in under 10 seconds.

¹ Adjustable thinking effort: you choose frugal, normal or thorough per chat. With GPT-OSS 120B the chain of thought gets about 20 times longer from frugal to thorough. Through the API this is the reasoning_effort field; which models evaluate it is listed by GET /api/v1/models in the effort_levels field (Chat Completions).

² GLM 5.3: its placing for programming rests on comparison figures from the vendor (Z.ai). We verified the function, not the ranking. The licence is its own, MIT-like one: full rights to use, modify and redistribute. Its single condition only applies above USD 10bn annual revenue and does not apply to merely forwarding requests to externally hosted models.

³ Context in partner operation: SovrGPT plans with 131k tokens, which is also the context_window value in GET /v1/models. According to their providers some of these models can handle more; we have not yet verified that with longer inputs.

⁴ Qwen Community License 1.0: Qwen 3.8 Flash is the only entry of the Qwen 3.8 line not under Apache 2.0. The licence requires a separate licence from Alibaba if you run a model-as-a-service business yourself. It explicitly excludes merely relaying requests to models hosted by third parties, which is exactly what this entry does.

⁵ Licence unverified: we have not read the licence file of these weights ourselves. Please talk to us before commercial use with an intent to redistribute.

⁶ Index = Artificial Analysis Intelligence Index, as of 2026-08-08. It measures these models in one shared test setup, which makes the values comparable with each other. "Not listed" means the source does not carry the model, and we do not estimate a value. The index is based on English-language tests and says nothing about quality in German.

⁷ Legacy: the model stays fully usable and addressable through the API, but it is no longer the recommended choice and will phase out in the medium term. There is no fixed shutdown date; existing chats and integrations continue unchanged.

When a model is missing: the org selection

Owners and admins of an organisation decide under Settings → Models which models their team sees. That keeps the model picker manageable. If a model is listed here but missing for you, this is the most likely reason.

The selection applies server-side, so it also covers the API and MCP clients. Running chats are unaffected: a chat that is already pinned to a model keeps running on it. Details: Orgs & teams.

Bringing your own models (BYO)

If the catalogue is not enough, an organisation can register its own OpenAI-compatible endpoints under Settings → Providers, with its own address, model name and display name. Typical cases are a self-operated vLLM in your own data centre, a domain model from a service provider or an Azure OpenAI deployment. The entry appears only in that organisation's model picker, marked "⚠ external".

What applies:

• Data path: requests go directly to the registered address. The terms of that endpoint apply; our commitments about the data path do not cover it. That is what the "external" marker shows.
• Tools and thinking control stay off. We have not verified a third-party endpoint, and a tool call to a model without a matching parser fails hard.
• Availability and waiting time depend on your endpoint alone.
• Addresses: https only. SovrGPT rejects addresses on local or private networks and raw IP addresses; an endpoint needs a DNS name with a valid certificate. SovrGPT does not follow redirects: if your endpoint redirects, you get HTTP 502 with the type upstream_redirect_refused. Enter the final address instead.
• Keys: a stored API key is saved encrypted (AES-256-GCM) and never leaves the server. Endpoints behind an IP allowlist work without any key.

You can pause or delete an entry at any time, including after a plan change that no longer includes creating new entries. Chats bound to it then answer with the default model again. Through the API, GET /v1/models lists these entries with owned_by: "org" and the ID org-model:….

Coding in the IDE

The coding model is mainly meant for agentic coding through IDE plugins (Kilo Code, Cline, OpenCode and others). They all speak the OpenAI-compatible API:

• Base URL: https://sovrgpt.com/api/v1
• API key: create one under /settings/api-keys (scope chat) and use it as a bearer token.
• Model ID: qwen3.6-35b-a3b, self-operated, 131k context. Without a cold start: stackit-qwen3.8-27b or tensorx-kimi-k2.7-code.

The full setup guide for every tool is in Coding in the IDE.

Self-operated models need a cold start of several minutes after a pause. Set your plugin's timeout to at least 900 seconds or use a model without a cold start, see Cold start and timeout.

Kilo Code (VS Code, JetBrains, command line): Settings → Providers → Custom provider · Provider API OpenAI Compatible · Base URL https://sovrgpt.com/api/v1 · API key sov_… · pick a model from the list, e.g. qwen3.6-35b-a3b. All fields are listed under Kilo Code.

Cline (VS Code): API Provider → OpenAI Compatible · Base URL https://sovrgpt.com/api/v1 · API Key sov_… · Model ID qwen3.6-35b-a3b · Context Window 131072 (must be set by hand).

OpenCode (opencode.json):

{
  "provider": {
    "sovrgpt": {
      "npm": "@ai-sdk/openai-compatible",
      "options": { "baseURL": "https://sovrgpt.com/api/v1", "apiKey": "sov_..." },
      "models": {
        "qwen3.6-35b-a3b": {
          "name": "SovrGPT Coder mini",
          "limit": { "context": 131072, "output": 8192 }
        }
      }
    }
  }
}

limit tells OpenCode the context window; for a provider you add yourself it does not know it otherwise.

Thinking effort per request (reasoning_effort)

Which models evaluate the control is stated by the effort_levels field of GET /v1/models. An empty array means no effective control. Query it per model: the levels are measured individually, and related models behave differently.

You steer it through the top-level field reasoning_effort (in the Python SDK via extra_body) with the levels low, medium and high. SovrGPT translates what each level becomes at the respective provider, such as a token cap or a different scale.

curl https://sovrgpt.com/api/v1/chat/completions \
  -H "Authorization: Bearer $SOVR_KEY" -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.6-35b-a3b",
    "messages": [{ "role": "user", "content": "Refactor this module and explain the steps." }],
    "reasoning_effort": "high",
    "max_tokens": 8192
  }'

With high, raise max_tokens as well, because the chain of thought counts against the answer budget. With qwen3.6-35b-a3b, an unlimited thinking budget and 2,000 max_tokens left the answer text empty.

For full control, pass chat_template_kwargs directly, for example {"thinking": true, "reasoning_effort": "max"}. It is forwarded unchanged and takes precedence over reasoning_effort. Details: POST /v1/chat/completions.

Understanding cold starts

• Warm: the model is loaded on a GPU, and the answer usually starts in under 10 seconds.
• Cold: a server has to start, load the image and fetch the model weights. Measured, this takes 2 to 10 minutes depending on the model (values in the tables above).
• Partner-operated: no cold start.

The self-operated servers shut down during pauses, and no compute costs accrue while idle. In return, the first request after a longer pause is slow. For scheduled production jobs we recommend sending a short warm-up request first. If you cannot wait at all, use a partner-operated model.

Tools in chat

• Web search through the Brave Search API (Brave Software, Inc., USA). According to Brave, search logs are kept for up to 90 days; Brave offers search without logs only on its Enterprise plan. In chat it can be switched on per conversation and is off by default; for the Excel add-in see there. Details under Data protection.

  The same switch lets the model open and read a page (fetch_url). That request goes from our servers straight to the site, without Brave. Owners and admins decide how SovrGPT identifies itself (Organisation → Page fetch): discreetly as an ordinary browser (the default; pages with blanket bot blocks are then more likely to serve content) or by name (SovrGPT names itself and rooom AG). "Discreet" only changes the headers, not our IP address or technical characteristics of the connection. Every change is recorded in the audit log.

• Image generation: Z-Image-Turbo (fast) or, in Pro mode, Qwen-Image 2512 (highest quality, strong at text inside the image), self-operated on rented GPUs in EU data centres. All image and video models are listed in the table below.

• Image editing: with image generation switched on in chat, the assistant edits one to three of your own images on instruction, i.e. attached images or earlier results from the same chat. For example, it can change colours, the background or objects, or combine several sources into one scene. The result is about 1 megapixel, and the aspect ratio comes from the last source. The model often renders text inside the image incorrectly. Through the MCP server the tool is called sovrgpt_edit_image.

• Video generation: LTX-Video 0.9.8 13B distilled, clips of 3–8 s at 768×512 and 24 fps, without sound. A clip is made from a description, from a start image (an attached PNG or JPEG up to 20 MB or an earlier image result from the same chat; ideally 3:2 landscape, other formats get distorted) or as a continuation of one of your own videos from the same chat: the new clip starts with its last frames and carries it on; it is a separate clip, not a merged video. Each job takes at most one source. The video generation terms of use apply.

• Marketplace connectors: over 20 preconfigured MCP servers (Marketplace).

Tools only work with models whose endpoint supports them. For other models SovrGPT locks the switches in the composer so that no answer fails because of them. Every model in the catalogue can use tools except llama-3.1-8b. In automatic mode that model calls no tools and answers in prose. In one test it even claimed a web search and invented a result. That is why the switch stays locked there.

Image and video models

All of them are self-operated on rented GPUs in EU data centres (and additionally in Iceland or Norway, EEA). After a pause a server has to start first; the last column shows how long that takes. In chat you switch generation on in the composer; through the API and the MCP server you pick the model by its model ID.

image-pro and Qwen-Image: Pro mode in chat now generates images with Qwen-Image 2512 where it is available, otherwise with FLUX.2 [klein] 4B. Through the API and MCP you pick Qwen-Image explicitly with image-qwen. The ID image-pro stands for FLUX.2 [klein] 4B; if FLUX is not ready, Qwen-Image handles the job.

video-pro generates clips of 3–10 s with an AI-generated audio track, at 960×544 or 1920×1088. The model is not enabled yet. Its licence contains use restrictions, among them no military applications; we will give the details at launch.

Images in chat

Attach an image to your message with the paper clip or by dropping it into the window.

If the selected model cannot look at images, SovrGPT says so above the input field before you send. The assistant also points it out in its reply and does not guess what the image shows. The image stays in the transcript; switching models is enough.

Putting an image on a slide works with any model. The model only passes the file name to export_pptx, and the file is inserted on the server. SovrGPT rejects oversized images; there is no automatic downscaling. Upload a smaller version instead.

None of these checks apply to the OpenAI-compatible API (/v1/chat/completions). That route forwards the request body unchanged: an image_url part reaches the model with no capability check, no size limit and no notice. If the model cannot read images, the request fails at the endpoint. We have not tested this path separately. For images through the API, pick a vision-capable model (accepts_vision: true).

Tokens, quotas and limits

• Standard plan: fair use, no hard token limit.
• Enterprise plan: terms by arrangement in a separate contract.
• Billing is per organisation, not per user.

Logged in, you see the pricing tiers under /settings/usage; your contract with rooom AG is binding.