Skip to main content

OVHcloud AI Endpoints Provider Guide

OVHcloud AI Endpoints is a Tier-2 catalog provider: OpenAI-wire-compatible with no behavioural quirks, so its entire integration is one JSON file (src/lib/providers/catalog/ovhcloud.json) rather than hand-written code. That file is the source of truth for everything on this page.

Verification status: this entry is docs- and roster-verified, not yet live-verified. It was built entirely from OVHcloud's own public documentation pages and one unauthenticated GET /v1/models call — no account was created, no key was issued, and no authenticated or POST request was ever sent. evidence.liveMatrix is null until someone runs the credentialed matrix (see Live verification below).

Key Facts​

  • Provider id: ovhcloud
  • Protocol: OpenAI-compatible (/chat/completions) — OVHcloud's own docs state the LLM APIs are "compatible with the OpenAI specifications"
  • Base URL: https://oai.endpoints.kepler.ai.cloud.ovh.net/v1
  • Default model: gpt-oss-120b
  • Models in catalog: 14 (chat-completions models only — see Out of scope)
  • Streaming: documented (the vendor's own gpt-oss-120b catalog card lists "Streaming" under Supported Features)
  • Tool calling: model-dependent — documented per-model on OVHcloud's catalog page (Yes for gpt-oss-120b, gpt-oss-20b, the four Qwen3.x visual models and Meta-Llama-3_3-70B-Instruct; No for Qwen2.5-VL-72B-Instruct)
  • Structured output: documented — OVHcloud's capabilities guide describes both JSON Mode (json_object) and JSON-Schema Structured Outputs, and the gpt-oss-120b catalog card lists json_object and json_schema under Output Formats
  • Tools + structured output together: not asserted. No combined probe was possible without credentials, so structuredOutputWithTools is false and toolsWithStreaming is conservatively false too — both features are independently documented, but nothing confirms they compose correctly in one request
  • Embeddings: not wired (OVHcloud does serve embedding models, but as separate endpoints outside this chat-completions integration's scope)
  • Billing: free-with-card — several individual models are billed at €0, but OVHcloud's own docs state a project in "Discovery mode" (no payment method attached) cannot use the API at all
  • Key format: none declared
  • Rate limits: documented as 400 requests/minute per project per model with a key; 429 on overage

Quick Start​

1. Get an API key​

  1. Visit: https://www.ovhcloud.com/en/public-cloud/ai-endpoints/ and click Start now, which signs you in to (or creates) an OVHcloud Public Cloud manager account
  2. In the manager, open Public Cloud > your project > AI & Machine Learning

    AI Endpoints, and add a payment method to the project — OVHcloud states that projects in Discovery mode (no payment method) cannot use the service

  3. Create an AI Endpoints API access token in that section; it is shown only once, so store it yourself, and note it can be given an expiration date and revoked
  4. Set OVH_AI_ENDPOINTS_ACCESS_TOKEN in your .env file (this is the exact variable name OVHcloud's own code samples use)

2. Configure​

export OVH_AI_ENDPOINTS_ACCESS_TOKEN=your-access-token
export OVHCLOUD_MODEL=gpt-oss-120b      # optional — overrides the default model
export OVHCLOUD_BASE_URL=https://oai.endpoints.kepler.ai.cloud.ovh.net/v1   # optional — self-hosted or proxy

3. Use it​

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  input: { text: "Explain context windows in one paragraph." },
  provider: "ovhcloud",
  model: "gpt-oss-120b",
});

console.log(result.content);
# CLI
npx @juspay/neurolink generate "Hello" --provider ovhcloud

Per-request credentials work as they do for every provider:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "ovhcloud",
  credentials: {
    ovhcloud: { apiKey: process.env.OVH_AI_ENDPOINTS_ACCESS_TOKEN },
  },
});

Models​

ModelContextVisionTools$/M in · outNotes
gpt-oss-120b ⭐131Knoyes$0.09 / $0.47Recommended — the vendor's own example model; docs confirm streaming, tools, reasoning, json_object/json_schema
gpt-oss-20b131Knoyes$0.05 / $0.18Smaller gpt-oss sibling; roster-listed, catalog-page confirmed; fallback
Qwen3.6-27B262Kyesyes$0.47 / $3.19Visual LLM (New); catalog-page confirmed; fallback
Qwen3.8-27B262Kyesyes$0.47 / $3.19Visual LLM (New), newer sibling of 3.6; catalog-page confirmed
Qwen3.5-397B-A17B262Kyesyes$0.71 / $4.25Largest MoE in the roster (397B params, ~17B active); catalog-page confirmed
Qwen3.5-9B262Kyesyes$0.12 / $0.18Smallest of the Qwen3.x visual family; catalog-page confirmed
Meta-Llama-3_3-70B-Instruct131Knoyes$0.74 / $0.74Plain non-reasoning instruct model; catalog-page confirmed
Qwen2.5-VL-72B-Instruct32Kyesno$1.01 / $1.01Dedicated vision-language model; models.visionModel. No tools per catalog page
Qwen3-Coder-30B-A3B-Instruct262Knon/a$0.07 / $0.26Live roster only — not yet on the marketing catalog page as of 2026-09-28
Mistral-Small-3.2-24B-Instruct-2506131Knon/a$0.10 / $0.31Live roster only — not yet on the marketing catalog page as of 2026-09-28
Mistral-7B-Instruct-v0.365Knon/a$0.11 / $0.11Live roster only — not yet on the marketing catalog page as of 2026-09-28
Mistral-Nemo-Instruct-240765Knon/a$0.14 / $0.14Live roster only — not yet on the marketing catalog page as of 2026-09-28
Qwen3Guard-Gen-8B32KnonofreeModeration/safety classifier, Beta — excluded from fallbacks/topModels
Qwen3Guard-Gen-0.6B32KnonofreeSmaller moderation/safety classifier, Beta — excluded from fallbacks/topModels

Context window and output cap per model are the live, unauthenticated GET /v1/models response's own context_length / max_completion_tokens fields (retrieved 2026-09-28). Pricing is that same response's per-token USD figures converted to $/MTok, cross-checked against the EUR prices on OVHcloud's public catalog page where a model appears there.

Fallback order when the default is unavailable: gpt-oss-20b → Meta-Llama-3_3-70B-Instruct → Qwen3.6-27B → Mistral-Small-3.2-24B-Instruct-2506 → Qwen3.5-9B.

Out-of-scope (non-chat) endpoints​

OVHcloud AI Endpoints also serves speech-to-text (whisper-large-v3, whisper-large-v3-turbo), text-to-speech (four NVIDIA Riva voices), image generation (stable-diffusion-xl-base-v10) and embeddings (bge-m3, bge-multilingual-gemma2, Qwen3-Embedding-8B) on the same platform and base URL prefix. These are separate, non-chat-completions API shapes and are not part of this Tier-2 catalog entry, so capabilities.embeddings is false here even though the vendor documents embedding models elsewhere on the same platform.

Verification status​

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This is what the catalog records for OVHcloud AI Endpoints — deliberately credential-free:

ProbeResult
Rosterunauthenticated GET /v1/models, HTTP 200, 2026-09-28 — 24 models returned; no key used
Auth rejectionnot probed — would require an authenticated request, which this credential-free onboarding does not send. The errorRules entry for authentication failures is taken verbatim from OVHcloud's written docs ("403 Forbidden: Authentication Failed"), not from a live probe
Live capability sweepnot run (evidence.liveMatrix is null). capabilities.tools, .streaming, .structuredOutput and .thinking are set from OVHcloud's own getting-started guide, capabilities guide, and the gpt-oss-120b catalog card's "Supported Features" / "Output Formats" lists — not from a live request of any kind

Because no live request was ever made, treat every capability flag here as doc-asserted, not wire-proven until a credentialed pass fills in evidence.liveMatrix per Tier-2 onboarding: Live verification.

Troubleshooting​

SymptomCauseFix
Requests fail immediatelyProject has no payment method attached ("Discovery mode")Add a payment method to the Public Cloud project in the OVHcloud manager before creating/using an access token
Invalid or revoked ... tokenOVH_AI_ENDPOINTS_ACCESS_TOKEN unset, wrong, or revokedCheck/recreate the token under AI & Machine Learning > AI Endpoints in the OVHcloud manager
Frequent 429sAuthenticated limit is 400 requests/minute per project per modelPace requests and/or request a higher limit from OVHcloud support
Model not foundThe roster changed since 2026-09-28Pick a current id from the unauthenticated GET /v1/models roster
Tool calls silently unsupportedcapabilities.tools is model-dependentUse a model the catalog page marks Function calling: Yes (e.g. gpt-oss-120b, not Qwen2.5-VL-72B-Instruct)

See also​