One API, multiple capabilities. Chat, embeddings, image generation, and speech. All running on GPUs in Groningen.

Models

GET /v1/models

Lists all available models. Returns model IDs and their capabilities.

Chat models

Text generation and conversation. Use with /v1/chat/completions. Both chat models support streaming and tool calling.

gemma

Gemma 4 31B (dense). Our fast chat model and the strongest in Dutch: it won a blind Dutch quality evaluation against every other model we tested. Use it for conversations, assistants, summaries and everyday text work.

~60 tok/s
64K context
2x NVIDIA A40

Recommended for chat and most interactive use cases.

qwen

Qwen 3.8 27B (dense). Our flagship long-context model, running on our fastest GPU. Use it for large documents, dossiers, retrieval-heavy workloads and anything that needs more than 64K of context. Prompt caching makes repeated requests over the same large document return in about two seconds.

~45 tok/s
256K context
NVIDIA RTX PRO 6000 (Blackwell)

Recommended for long documents and large contexts. Thinking is off by default. Enable it per request with "chat_template_kwargs": {"enable_thinking": true}; it costs a few thousand extra tokens per answer.

Choosing between the two

Workload	Model
Chat, assistants, Dutch text	`gemma`
Summaries and extraction on documents under 64K tokens	`gemma`
Documents, dossiers and transcripts over 64K tokens	`qwen`
Repeated questions over the same large context	`qwen`

If one chat model is unavailable, requests are served by the other. The response then carries the header X-Douane-Fallback with the model that answered.

Deprecated model names

These names keep working for now, but new code should use gemma or qwen.

qwen-fast, qwen-deep, qwen38: aliases for qwen.
Qwen/Qwen3.6-35B-A3B-FP8: Qwen 3.6 has been replaced by Qwen 3.8. This name resolves to qwen until 6 October 2026, then it is removed.
Qwen/Qwen3.5-35B-A3B-FP8 and Qwen/Qwen3.5-27B-FP8: Qwen 3.5 is retired. These names currently resolve to gemma and will be removed. Switch to gemma for chat or qwen for long documents.
flux-dev: retired. Requests return an error pointing to flux-schnell.

Embedding models

Convert text to vectors for search, similarity, and RAG. Use with /v1/embeddings.

bge-m3

BGE-M3 multilingual embeddings. State-of-the-art for retrieval, supports 100+ languages including Dutch.

1024 dimensions
8K tokens max

Image generation

Generate images from text. Use with /v1/images/generations.

flux-schnell

FLUX Schnell. Fast generation (~2s per image) for rapid iteration and prototyping.

1024×1024
~2s

Speech to text

Plain transcription with /v1/audio/transcriptions (OpenAI-compatible, model whisper-1), or with speaker diarization on /v1/audio/diarize, and /v1/audio/diarize/jobs for long recordings (see Speech to text).

whisperx

WhisperX on Whisper large-v3. Transcribes audio and, on the diarize endpoint, identifies who said what. Supports Dutch and 90+ other languages.

MP3, WAV, FLAC
Speaker labels (diarize endpoint)

Model aliases

Use short aliases or full model names interchangeably.

Alias	Model
`gemma`	RedHatAI/gemma-4-31B-it-FP8-block
`gemma-4`	RedHatAI/gemma-4-31B-it-FP8-block
`qwen`	Qwen/Qwen3.8-27B-FP8
`qwen-fast`, `qwen-deep`, `qwen38` (deprecated)	Qwen/Qwen3.8-27B-FP8
`flux-schnell`	schnell
`diarize`	whisperx
`whisper-1`, `whisper-large-v3`	whisperx

Need a different model? We can deploy additional models on request. Contact us at support@appelon.ai