Available Models¶
Models change quickly
Our model lineup changes every few months: a model usually stays available for 3–4 months, at most about 6. New models and retirements are announced in the Release Notes and the Matrix room. If your work depends on one specific model, for example your thesis, read How long does a model stay available? first.
All models are available through the OpenAI-compatible API at https://aqueduct.ai.datalab.tuwien.ac.at/v1
and in Open WebUI.
Upcoming deprecations — 24 November 2026
kokoro, piper-thorsten and z-image-turbo are scheduled for deprecation on 24 November 2026.
Language Models¶
| Model ID | Context | Vision | Reasoning efforts | Limits | Status |
|---|---|---|---|---|---|
deepseek-v4-vision-exp-284b |
393,216 | low · high · max | 1× | Stable | |
glm-5.3-744b-experimental |
262,144 | low · high · max | 1× | Experimental | |
qwen-3.8-27b |
262,144 | low · medium · xhigh | 2× | Stable |
Context is the maximum usable context window in tokens, including the output. Limits is the multiplier on your personal rate limits.
Model details¶
deepseek-v4-vision-exp-284b- DeepSeek-V4-Flash-Vision-Exp on Hugging Face. Our main model. 2 instances, each running on 4x NVIDIA RTX PRO 6000 Blackwell.
glm-5.3-744b-experimental- GLM-5.3 on Hugging Face. Experimental deployment: please only use it for coding or day-to-day tasks. It may be removed at short notice. 1 instance, running on 8x AMD Instinct MI300X.
qwen-3.8-27b- Qwen3.8-27B-FP8 on Hugging Face.
Fast, smaller model, also available under the aliases
visionandocr. 4 instances, each running on 1x NVIDIA H200.
Context and output tokens¶
The context lengths in the table are the maximum usable values: longer inputs are trimmed to that size, and output
tokens count toward the context window. Answers can be up to 65536 tokens long by default; set max_tokens to change
this per request.
Some clients (e.g. Pi, Continue) have their own maxTokens setting that cuts answers short. Set it to 65536 to use the
full output budget.
Reasoning¶
The language models can think (reason) before they answer. You can control this per request:
reasoning_effort: how much the model thinks, with the values from the Reasoning efforts columnenable_thinking:falseturns thinking off,trueturns it on, overriding the server default
With the OpenAI-compatible API, both go in chat_template_kwargs:
# client = OpenAI(...)
response = client.chat.completions.create(
model=model,
messages=messages,
extra_body={
"chat_template_kwargs": {
"enable_thinking": True, # overrides server default
"reasoning_effort": "medium",
},
},
)
Some clients (e.g. older versions of certain coding tools) don't pass reasoning_effort through. Aqueduct then returns
aqueduct.UnsupportedParamsError: update the client or use a model alias without reasoning. For all request-level
options, see reasoning outputs in
the vLLM documentation.
Text-to-Speech (TTS)¶
See the speech guide in the public Aqueduct documentation for usage.
| Model ID | Language | Voices | Sample |
|---|---|---|---|
voxtral-4b-tts |
9 languages | 20 preset voices | — |
voxtral-4b-tts supports English, French, Spanish, Portuguese, Italian, Dutch, German, Arabic and Hindi.
English and German voice presets:
| Language | Voice presets |
|---|---|
| English | neutral_male, neutral_female, casual_male, casual_female, cheerful_female |
| German | de_male, de_female |
Speech-to-Text (STT)¶
See the speech guide in the public Aqueduct documentation for usage.
| Model ID | Languages |
|---|---|
whisper-large |
99 languages, including English and German |
Image Generation¶
See the image generation guide in the public Aqueduct documentation for usage.
| Model ID |
|---|
qwen-image-2-1-turbo |
z-image-turbo |
qwen-image-2-1-turbo is the 8-step Turbo checkpoint for text-to-image generation. Aqueduct currently does not support
image-edit endpoints.
Listing models via the API¶
The API always returns the current lineup:
curl https://aqueduct.ai.datalab.tuwien.ac.at/v1/models \
-H "Authorization: Bearer $AQUEDUCT_API_KEY" | jq
How well each model supports specific API features (streaming, tool calling, structured output, …) is tested nightly, see the Model Compatibility report.