Open-weight models
Qwen. Mistral. MMS-TTS. Hosted.
Open-weight HuggingFace models via a typed async API. Ready on demand, flat monthly pricing.
The hosted model registry covers instruct, embedding, text-to-speech, image-to-text, depth estimation, segmentation, and eval model types. One API, one auth token, no per-token billing regardless of how many times you call any model.
Hosted model registry
Instruct (text generation)
Instruction-following and chat completion. OpenAI-compatible endpoint available. Qwen2.5 covers 1.5B, 7B, and 14B parameter variants.
Text and image embedding
Dense vector representations for search, clustering, and retrieval. CLIP serves both image and text embedding, enabling cross-modal similarity in a single pipeline.
Image to text
Vision-language model for captioning, structured description, and visual question answering. Accepts image references as input.
Text to speech
Language-specific synthesis via the Facebook MMS family. Welsh (mms-tts-cym) is a specific capability absent from most commercial TTS services.
Depth and segmentation
Monocular depth estimation and image segmentation for vision pipelines. Used in change detection, conformance checking, and spatial analysis.
Evaluation models
Text, image, and cross-modal scoring for pipeline quality assessment. Run any model against a labelled dataset using the same API.
Calling a model
The same API across all model types
Submit via the OpenAI-compatible endpoint for instruct models, or via the native async API for any model type. Binary outputs are returned as storage references, not inlined in the response.
Qwen instruct via OpenAI-compatible endpoint
curl https://api.marigold.run/v1/chat/completions \
-H "Authorization: Bearer your-api-key" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen/qwen2.5-7b-instruct",
"messages": [{"role": "user", "content": "Translate to Welsh: Good morning."}]
}'TTS via native async API
import requests, time
API = "https://api.marigold.run"
KEY = {"Authorization": "Bearer your-api-key"}
job = requests.post(f"{API}/infer", headers=KEY, json={
"model_type": "tts",
"model_name": "facebook/mms-tts-cym",
"text": "Bore da, sut ydych chi?"
}).json()
while True:
r = requests.get(f"{API}/infer/{job['job_id']}", headers=KEY).json()
if r["status"] == "complete":
print(r["output"]) # storage reference to audio file
break
time.sleep(0.5)Custom models
Request a model not in the registry
Any HuggingFace model compatible with the standard handler types can be onboarded. Custom model onboarding is available on the Pro tier.
Send the HuggingFace model ID and intended use case. If it fits an existing handler type, onboarding is typically a day's work.
Request a modelOpen weights. Private infrastructure.
Leave your email to be notified when access opens. Mention specific models and we will confirm registry availability.