Skip to main content

STACKIT Provider Guide

STACKIT is a Tier-2 catalog provider: its integration is one JSON file (src/lib/providers/catalog/stackit.json) rather than hand-written code.

Verification status: this entry is docs-verified only, not yet live-verified. STACKIT's GET /models needs a key (an unauthenticated GET /v1/models on 2026-09-29 answered HTTP 403), so the model ids come from STACKIT's public documentation and the roster has not been checked. No account was created and no API key was used to build it. evidence.liveMatrix is null until someone runs the live capability matrix with a real key (see Verification status below).

Key Facts​

  • Provider id: stackit
  • Protocol: OpenAI-compatible (/chat/completions) — the model rows on the Available Shared Models page show Specification "OpenAI-compatible" and the endpoints POST /chat/completions, POST /completions and GET /models. The sample reply shown under the POST /v1/chat/completions curl request on the Getting started page reads "object":"text_completion" and "choices":[{"index":0,"text":; no POST request was sent while building this entry
  • Base URL: https://api.openai-compat.model-serving.eu01.onstackit.cloud/v1
  • Default model: Qwen/Qwen3-VL-235B-A22B-Instruct-FP8
  • Models in catalog: 6 chat models, taken from the "Full Name:" headings of the Available Shared Models page (retrieved 2026-09-29)
  • Streaming: not declared (false); no streaming request was sent
  • Tool calling: model-dependent. The Features cell reads "Tool calling enabled" on the Qwen3-VL 235B and Llama 3.3 70B rows and "Tool calling & Reasoning enabled" on the other four catalogued rows. The tool-calling how-to is a LangChain example that prints tool_call_response.tool_calls
  • Tools while streaming: not declared (false)
  • Structured output: not declared (false). The release note of Dec 10, 2025 says of openai/gpt-oss-120b: "The model provides full chain-of-thought (CoT) and support Structured Outputs." The catalog's structuredOutput flag applies to the provider as a whole, so it stays false
  • Structured output + tools together: not declared (false) — no combined probe was run without credentials
  • Embeddings: not declared on this catalog entry. STACKIT does list two embedding models with a POST /embeddings endpoint (intfloat/e5-mistral-7b-instruct, Qwen/Qwen3-VL-Embedding-8B), but NeuroLink's generic ConfiguredOpenAICompatProvider does not implement embed()/embedMany(), so this flag tracks that, not the vendor's own API surface
  • Thinking: not declared (false); the per-model Features text is recorded in the table below
  • Billing: no-free-tier — see Billing below
  • Key format: none declared

Quick Start​

1. Get an API key​

  1. Visit: https://portal.stackit.cloud and log in to the customer portal — the Getting started page says "To enable AI Model Serving login to the Customer portal and click at the sidebar on the left on AI Model Serving." and lists a STACKIT customer account, a STACKIT user account and a STACKIT project as prerequisites
  2. Enable AI Model Serving, click Create token, enter a token name and optionally a lifetime in days, and confirm with "Order fee-based" ("To confirm click on Order fee-based."); the same page says "You can not retrieve the token again after closing the pane", so copy it when it is shown
  3. Billing as the vendor states it: the May 6, 2025 release note says "STACKIT AI Model Serving offers you easy pay-as-you-go access to proven GenAI models"; the product page offers "Register for a free trial access" — see Billing
  4. Set STACKIT_API_KEY in your .env file (the docs' examples read STACKIT_MODEL_SERVING_AUTH_TOKEN and STACKIT_MODEL_SERVING_API_KEY, which NeuroLink does not read)

2. Configure​

export STACKIT_API_KEY=your-auth-token
export STACKIT_MODEL=Qwen/Qwen3-VL-235B-A22B-Instruct-FP8   # optional — overrides the default model
export STACKIT_BASE_URL=https://api.openai-compat.model-serving.eu01.onstackit.cloud/v1   # optional — proxy or self-hosted gateway

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: "stackit",
  model: "Qwen/Qwen3-VL-235B-A22B-Instruct-FP8",
});

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

Per-request credentials:

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

Billing​

What the vendor's pages state, quoted as they appear:

  • Getting started, token creation: "To confirm click on Order fee-based." The Manage auth tokens page's token-creation steps say "Click on Order fee-based."
  • Release note, May 6, 2025: "STACKIT AI Model Serving offers you easy pay-as-you-go access to proven GenAI models"
  • Product page, advantages list: "Operate GenAI applications cost-effectively with our token-based billing."
  • Product page, heading "Register for a free trial access": "All registrations are manually checked by us. If you qualify, you will receive your personal token to use STACKIT AI Model Serving within 24 hours. The trial period ends automatically after 30 days."

The product page's "STACKIT AI Model Serving price table" (Region cell "Germany South") shows these text-model rows, quoted as the cells read:

SKUBilling"Price per hour" cell
Model Serving-llm-standard-input-EU01mio input token0.15000000000 €
Model Serving-llm-standard-output-EU01mio output token0.25000000000 €
Model Serving-llm-plus-input-EU01mio input token0.45000000000 €
Model Serving-llm-plus-output-EU01mio output token0.65000000000 €
Model Serving-llm-premium-input-EU01mio input token1.50000000000 €
Model Serving-llm-premium-output-EU01mio output token1.75000000000 €

The Available Shared Models page gives each chat model a Category cell (LLM-Premium, LLM-Plus or LLM-Standard), shown in the table below.

The entry records billingPolicy: no-free-tier because the catalog schema takes one of three values and the token-creation step says "To confirm click on Order fee-based."; the free trial above is a separate, manually checked registration. The entry leaves pricingPerMTok empty: the vendor's prices are in euro and are recorded here as the vendor shows them.

Models​

ModelCategoryContextMax outputVisionRPM / TPMNotes
Qwen/Qwen3-VL-235B-A22B-Instruct-FP8 ⭐LLM-Premium200K16384yes (10 images)30 / 350000NeuroLink default. Features "Tool calling enabled". The row's text: "Meet Qwen3-VL – the most powerful vision-language model in the Qwen series to date."
cortecs/Llama-3.3-70B-Instruct-FP8-DynamicLLM-Plus128K4096no80 / 200000Fallback. Features "Tool calling enabled". The model in the Getting started curl example ("In this guide we use Llama 3.3 70B.").
openai/gpt-oss-120bLLM-Plus131K8192no30 / 200000Fallback. Features "Tool calling & Reasoning enabled".
Qwen/Qwen3.8-27BLLM-Plus262K16384yes (10 images)45 / 350000Fallback. Features "Tool calling & Reasoning enabled"; Thinking budget "8192 Token".
google/gemma-4-31B-itLLM-Plus256K4096yes (10 images)80 / 200000Fallback. Features "Tool calling & Reasoning enabled".
openai/gpt-oss-20bLLM-Standard131K8192no80 / 200000Fallback. Features "Tool calling & Reasoning enabled". Row note: "We recommend configuring retries for tool calling to improve reliability."

The figures in the table are cells of the Available Shared Models page (retrieved 2026-09-29); the six rows show Type "Chat" and Status "Supported". The page writes context as "200K Token", "128K Token", "131K Token", "262K Token" and "256K Token"; the entry records those as 200000, 128000, 131000, 262000 and 256000, reading K as thousands. Vision comes from each row's Modalities cell ("Input are text and image and Output is text." for the three vision rows, "Input and output are text." for the others). RPM and TPM are the page's "RPM limit" and "TPM limit" cells; the rate-limits page says limits are applied per STACKIT project.

models.defaultContextWindow (128000) and models.defaultMaxOutputTokens (4096) are placeholders the vendor does not publish as general values; they apply to model ids outside the catalog and equal the Context length "128K Token" and Maximum Generation "4096 Token" cells of the Llama 3.3 70B row.

The status field of the six catalogued models holds production: the catalog schema requires a status (production, preview or retired) and the vendor's rows show Status "Supported".

Fallback order when the default is unavailable: cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic → openai/gpt-oss-120b → Qwen/Qwen3.8-27B → google/gemma-4-31B-it → openai/gpt-oss-20b. The runtime fallback model name the loader derives (fallbacks[1]) is openai/gpt-oss-120b.

Left out of the catalog:

  • Two chat rows show Status "Deprecated": Qwen/Qwen3.6-27B and google/gemma-3-27b-it. The release notes say "We kindly ask all customers to migrate their workloads to the new model Qwen/Qwen3.8-27B before December 08, 2026." and "We kindly ask all customers to migrate their workloads to the new model google/gemma-4-31B-it before 14 October 2026."
  • Two embedding rows (intfloat/e5-mistral-7b-instruct, Qwen/Qwen3-VL-Embedding-8B), for the embeddings reason in Key Facts.

The Use the models page says "STACKIT AI Model Serving provides an OpenAI-compatible API, making it easy to integrate with existing tools and libraries." and "Important: The API is stateless and does not maintain session information." The Available Shared Models page gives a request body limit: "The maximum allowed request body size is 20 MiB."

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 currently records for STACKIT — docs-verified only, not live-verified:

ProbeResult
Rosterunauthenticated GET /v1/models answered HTTP 403 (content-length 0, no body) on 2026-09-29, so the roster was not read; the six ids come from the Available Shared Models page, retrieved 2026-09-29 — no key used
Billingvendor pages quoted in Billing, 2026-09-29 — no signup performed
Auth-failure shapeunauthenticated GETs of /v1/models, /v1/chat/completions and /v1 each answered HTTP 403 with content-length 0 and no body on 2026-09-29, so no error code was observable; no POST was sent. errorRules carry the two statuses the vendor's pages state: 404 (FAQ) and 429 (rate-limits page)
Chat response shapeThe sample reply shown under the Getting started page's POST /v1/chat/completions curl request reads "object":"text_completion" and "choices":[{"index":0,"text":; no POST request was sent, so no chat-completions reply was read directly
Tools / structured outputper-model Features cells and the tool-calling how-to are recorded above; neither was exercised live, and no combined tools+schema request was sent — structuredOutputWithTools stays false
Live capability sweepnot run. evidence.liveMatrix is null. Before treating this provider as production-ready, run npx tsx test/continuous-test-suite-provider-matrix.ts --provider=stackit with a real key and record the result.

Do not treat this entry as equivalent to a live-verified Tier-2 provider (e.g. FriendliAI, Novita AI) until that live matrix has been run and evidence.liveMatrix is filled in.

Troubleshooting​

Each row is what a STACKIT page says, with the page named.

SymptomWhat the vendor's page saysPage
HTTP 404"Most likely this is due to an incorrect “model” parameter in the request body." and "This means a chat model cannot be used to compute embeddings, or vice versa."FAQ
HTTP 429"Once you hit a rate limit (RPM or TPM)" the API responds with 429 Too Many Requests; a heading on the page reads "Implement exponential backoff with jitter"; the page lists the headers x-ratelimit-remaining-requests and x-ratelimit-remaining-tokensRate limits
Token does not workUnder "Why does my authentication token (aka API-key) not work?": "API base URL: https://api.openai-compat.model-serving.eu01.onstackit.cloud/v1" and "API key / Authentication token / Secret key: STACKIT AI Model Serving Auth Token"FAQ
Output truncated"If it is set to "length", the generation was stopped due to the token cap." (the finish_reason field)FAQ
LengthFinishReasonError with structured output"This problem can be solved by adjusting the frequency_penalty parameter. A value of 0.7 or higher has proven to be sufficient."FAQ
Llama 3.3 70B produces a hallucinated tool callUnder known issues, for cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic: "To avoid hallucinated tool calls, do not send the tools parameter with an empty list."FAQ
Tool calling on GPT-OSS 20B"We recommend configuring retries for tool calling to improve reliability."Available Shared Models

See also​