# Aqueduct Release Notes

<!-- Add each entry here in English and in release-notes.de.md in German. -->

## 2026-11-10

**New models for image and speech generation**

Two new models are now available through Aqueduct's OpenAI-compatible API:

* [Qwen Image 2.1 Turbo](https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo) (`qwen-image-2-1-turbo`) generates images from text prompts.
* [Voxtral 4B TTS](https://huggingface.co/mistralai/Voxtral-4B-TTS-2603) (`voxtral-4b-tts`) turns text into expressive speech in nine languages.

Aqueduct currently supports text-to-image generation only; image-edit endpoints are not supported.

`kokoro`, `piper-thorsten` and `z-image-turbo` are scheduled for deprecation on **24 November 2026**. See [Available Models](https://datalab.tuwien.ac.at/aiml/aqueduct/models/) for the current lineup.

## 2026-10-14

**Qwen 3.8 27B is now available**

**Qwen 3.8 27B** is now available in Aqueduct as `qwen-3.8-27b`. It is deployed on our H200 node with four instances, scaling up our small-model offering with more capacity.

**Model details:**

* [Qwen/Qwen3.8-27B-FP8](https://huggingface.co/Qwen/Qwen3.8-27B-FP8) (FP8)
* 262,144 tokens context length
* supports `low`, `medium` and `xhigh` reasoning efforts
* supports image inputs

**Aliases**

The model is also available under the aliases `vision` and `ocr`, and it has 2x the rate limits of our other models.

**Qwen 3.6 35B deprecation**

Our previous small model, `qwen-3.6-35b`, will be deprecated after 27 October. Depending on demand, we may scale it down to fewer instances earlier than that.

## 2026-09-30

**Qwen 3.5 retired**

As announced, Qwen 3.5 has been retired today and is no longer available in Aqueduct.

**DeepSeek V4 Flash with vision support**

An updated version of DeepSeek V4 Flash with vision support is available from today as `deepseek-v4-vision-exp-284b`. This means our main model tier supports image inputs again.

**Model details:**

* [deepseek-ai/DeepSeek-V4-Flash-Vision-Exp](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp)
* supports image inputs
* 393.216 tokens context length
* supports `low`, `high` and `max` reasoning efforts

Our `main` and `coding` aliases now point to this new model.

If you use these aliases, you don't need to change anything. If you reference `deepseek-v4-flash-284b` directly, please switch to `deepseek-v4-vision-exp-284b`.

**Transition period:**

* The new model is technically not exactly the same model as our current `deepseek-v4-flash-284b`. That is why we will run both models in parallel for one week, **until 7 October**.
* Please use this time to test your use cases with the new model. If you notice any differences in behaviour or run into issues, let us know.
* After this week, `deepseek-v4-flash-284b` will be removed.

We will share more information about the next model releases when we remove the old DeepSeek V4 Flash.

## 2026-09-17

Our **Qwen 3.5** Model will be retired **on 30 September** .

**Replacement models**

DeepSeek-V4-Flash is now supported by the RTX 6000s GPUs. This allows us to expand its deployment on that node and use it as the replacement for Qwen 3.5.

Moving the current DeepSeek model off the H200 node will also free up capacity there. We are considering **Qwen3.8-27B** for that node. As our most powerful node, the H200 could host four instances and help us scale our small-model offering.

**Possible future releases**

We are monitoring two candidates for a later upgrade on the RTX 6000 node:

- **DeepSeek-V4-Flash-Vision-Exp** depends on the fix in [flashinfer#4802](https://github.com/flashinfer-ai/flashinfer/pull/4802), which is included in [flashinfer v0.7.0rc3](https://github.com/flashinfer-ai/flashinfer/releases/tag/v0.7.0rc3), released yesterday. We still need to verify that it runs on our hardware.
- **GLM-5.3-Flash** depends on the same fix, as well as [flashinfer#4947](https://github.com/flashinfer-ai/flashinfer/pull/4947) and either [vLLM#55277](https://github.com/vllm-project/vllm/pull/55277) or [vLLM#53969](https://github.com/vllm-project/vllm/pull/53969).

Either model could be a follow-up option, depending on compatibility and testing. There is no confirmed upgrade date yet.

Further, we are also still working with the ASC team to resolve the issues with the AMD node. Once it is operational, **DeepSeek-V4.1-Flash** or **GLM-5.3** are possible deployment candidates, subject to compatibility and testing.

## 2026-08-07

**DeepSeek V4 Flash** is now available in Aqueduct as `deepseek-v4-flash-284b`.

It shows strong [benchmark performance](https://artificialanalysis.ai/?models=minimax-m3%2Cdeepseek-v4-flash%2Cglm-5-2%2Ckimi-k3&intelligence=coding-index#intelligence-tabs), especially for its size. That combination makes it a very strong candidate right now.

**One thing to note:** DeepSeek does not support image inputs. This means we will lose that modality for our large model. Image input will still be possible via our Qwen 3.6 model. Feel free to provide us with feedback if this breaks any of your current use cases.

**Model details:**

* [deepseek-ai/DeepSeek-V4-Flash-0731](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731)
* original model without any extra quantization applied
* 393,216 tokens context length
* supports `high` and `max` reasoning efforts

Rollout plan:

* DeepSeek V4 Flash now runs in parallel on our H200s while Qwen 3.5 continues on the RTX 6000 Pros
* Currently, there are upstream issues in vLLM w.r.t. the deployment of DeepSeek V4 on RTX Pro 6000 Blackwell.
* Qwen 3.5 will eventually be fully deprecated once DeepSeek runs on RTX Pro 6000 Blackwell. DeepSeek will then inherit the "main" model alias.
* We will update you with a concrete timeline as soon as we have a clear picture, but not earlier than 21st of August.