OVH AI Endpoints

OVH AI Endpoints Provider

This project is a Dify model provider plugin for OVH AI Endpoints.

This provider is configured for OVH AI Endpoints with a predefined model catalog.

What users need to configure

Only credentials:

• api_key (required)

No endpoint URL is required in UI. The OVH OpenAI-compatible base URL is injected automatically for LLM, moderation, embeddings, and speech-to-text. TTS uses OVH native per-model endpoints.

OVH references:

• AI Endpoints overview: https://www.ovhcloud.com/en/public-cloud/ai-endpoints/
• AI Endpoints catalog: https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/

OVH AI Endpoints Models Matrix

Reference date: 2026-03-31.

Workflow code-node profiles

• LLM_STANDARD: temperature, top_p, max_tokens, structured output controlled by the native Dify LLM/Agent blocks.
• LLM_REASONING: LLM_STANDARD + reasoning_effort (low|medium|high) and optional hidden reasoning output handling.
• LLM_VISION: LLM_STANDARD with multimodal messages payloads (text + image_url) and, when exposed by the model, reasoning_effort.
• MODERATION_GUARD: moderation/safety classification, boolean blocked/safe output.
• EMBEDDINGS: endpoint /v1/embeddings, body {model, input} where input is str|list[str], output vectors.
• SPEECH2TEXT: endpoint /v1/audio/transcriptions, multipart body with file, model, optional language, prompt, response_format.
• TTS: OVH native endpoint https://{model}.endpoints.kepler.ai.cloud.ovh.net/api/v1/tts/text_to_audio, body {encoding, language_code, sample_rate_hz, text, voice_name}, output WAV audio bytes converted to MP3 before returning to Dify.
• IMAGE_GENERATION: endpoint /v1/images/generations, body {model, prompt, size, response_format}.

Per-model mapping

Model	OVH category	Dify type in this provider	I/O summary	Profile
gpt-oss-120b	Reasoning LLM	llm	text/json out, tool calling, reasoning	LLM_REASONING
gpt-oss-20b	Reasoning LLM	llm	text/json out, tool calling, reasoning	LLM_REASONING
Qwen3.6-27B	Visual LLM	llm	multimodal in, text/json out, tool calling, reasoning	LLM_VISION
Qwen3.5-397B-A17B	Visual LLM	llm	multimodal in, text/json out, tool calling, reasoning	LLM_VISION
Qwen3.5-9B	Visual LLM	llm	multimodal in, text/json out, tool calling, reasoning	LLM_VISION
Mistral-Small-3.2-24B-Instruct-2506	Visual LLM	llm	multimodal in, text/json out, tool calling	LLM_VISION
Qwen2.5-VL-72B-Instruct	Visual LLM	llm	multimodal in, text/json out	LLM_VISION
Meta-Llama-3_3-70B-Instruct	LLM	llm	text/json out, tool calling	LLM_STANDARD
Mistral-Nemo-Instruct-2407	LLM	llm	text/json out, tool calling	LLM_STANDARD
Qwen3Guard-Gen-8B	LLM Guard	moderation	text in, blocked/safe boolean out	MODERATION_GUARD
Qwen3Guard-Gen-0.6B	LLM Guard	moderation	text in, blocked/safe boolean out	MODERATION_GUARD
Qwen3-Embedding-8B	Embeddings	text-embedding	text in, vectors out	EMBEDDINGS
bge-multilingual-gemma2	Embeddings	text-embedding	text in, vectors out	EMBEDDINGS
bge-m3	Embeddings	text-embedding	text in, vectors out	EMBEDDINGS
whisper-large-v3	Automatic Speech Recognition	speech2text	audio in, transcript out	SPEECH2TEXT
whisper-large-v3-turbo	Automatic Speech Recognition	speech2text	audio in, transcript out	SPEECH2TEXT
nvr-tts-en-us	Text to Speech	tts	text in, MP3 audio out	TTS
nvr-tts-de-de	Text to Speech	tts	text in, MP3 audio out	TTS
nvr-tts-it-it	Text to Speech	tts	text in, MP3 audio out	TTS
nvr-tts-es-es	Text to Speech	tts	text in, MP3 audio out	TTS
stable-diffusion-xl-base-v10	Image Generation	not exposed as Dify model type	prompt in, image out	IMAGE_GENERATION

Notes

• For TTS, OVH currently returns WAV from its native endpoint and the plugin converts it to MP3 before returning it to Dify.
• Chrome is known to play the current TTS path correctly. Firefox may fail in Dify UI depending on browser MediaSource support for the MIME used by the frontend.

Context windows from OVH catalog

• gpt-oss-120b: 131K
• gpt-oss-20b: 131K
• Qwen3.6-27B: 262K
• Qwen3.5-397B-A17B: 262K
• Qwen3.5-9B: 262K
• Mistral-Small-3.2-24B-Instruct-2506: 128K
• Meta-Llama-3_3-70B-Instruct: 131K
• Qwen2.5-VL-72B-Instruct: 32K
• Mistral-Nemo-Instruct-2407: 118K
• Qwen3Guard-Gen-8B: 32K
• Qwen3Guard-Gen-0.6B: 32K

Submission metadata

• Repository: https://github.com/rungeard/dify_ovh_ai_endpoints
• Contact: https://github.com/rungeard/dify_ovh_ai_endpoints/issues
• Privacy policy: ./PRIVACY.md