--- title: "GWDG KISSKI Chat AI and SAIA free inference" description: "No-cost research and development inference across 16 models; not unrestricted production use." canonical_url: "https://freeinferencing.com/provider/gwdg_saia/" md_url: "https://freeinferencing.com/provider/gwdg_saia.md" last_updated: "2026-09-19" --- # GWDG KISSKI Chat AI and SAIA > No-cost research and development inference across 16 models; not unrestricted production use. ## Classification - **Directory:** Current - **Free access:** Research / development - **Status:** Restricted access - **Confidence:** High - **Payment card:** Unknown - **Account:** Required - **Equivalent paid value:** Not quantifiable - **API endpoint:** `https://chat-ai.academiccloud.de/v1` ## Models mentioned - `Apertus-70B-Instruct-2509` - `DeepSeek-V4-Flash-0731` - `Devstral-2-123B-Instruct-2512` - `Gemma-4-31B-Instruct` - `GLM-4.7` - `Llama-3.1-8B-Instruct` - `Mistral-Medium-3.5-128B` - `GPT-OSS-120B` - `Qwen-3-30B-A3B-Instruct-2507` - `Qwen-3-Coder-Next` - `Qwen-3-Omni-30B-A3B-Instruct` - `Qwen-3.5-122B-A10B` - `Qwen-3.5-397B-A17B` - `Qwen-3.6-35B-A3B` - `Qwen-3.8-27B` - `E5-Mistral-7B-Instruct` ## Limits and terms ```yaml requests_per_minute: 10 requests_per_hour: 200 requests_per_day: 400 requests_per_month: 3000 ``` ## What happens to your prompts? **Private.** For qualifying GWDG-hosted routes, prompts and responses are not persistently stored or used for AI training; metadata-only logging and narrow operational exceptions remain. Privacy is audited separately from price. Review the current governing terms before sending sensitive or regulated data. ## Data governance and agreements ```yaml data_governance: review_status: reviewed plan_scope: GWDG-hosted open-weight Chat AI/SAIA routes; external commercial routes are excluded. prompt_retention: Prompts are processed in working memory and not persistently stored; optional uploads may differ. response_retention: Core service responses are not persistently stored. ordinary_logging: Metadata can include username, timestamp, model, request/response length, and token information; some usage data can remain for one year. model_training: Prompts and responses are not used to train AI models. product_improvement: No content-based improvement use is stated for qualifying routes. human_or_operator_access: Ordinary content is not persistently available; transient support/security access is not fully enumerated. subprocessors_and_routing: Qualifying models run on GWDG infrastructure; optional tools and commercial routes have separate terms. deletion_controls: Core content is not persistently stored; account, metadata, upload, and optional-feature controls vary. ``` ## Eligibility ```yaml account_required: true eligible_users: Students, researchers, and affiliates of participating universities or research institutions identity: Academic ID or federated login ``` ## API compatibility and modalities - **Compatibility:** OpenAI Compatible - **Modalities:** Text generation - **Geography:** Germany ## Before you build with GWDG KISSKI Chat AI and SAIA ### Read the classification narrowly This record describes the bounded offer supported by evidence on 2026-09-19; it does not guarantee permanence, production suitability, uptime, latency, model quality, or access from every account and region. “Research / development” should be interpreted together with the Current directory placement, High confidence, Unknown payment-card status, and the plan-specific sources below. ### Resolve the live model route The snapshot records 16 model IDs. Match the exact ID against the provider's current catalog before using it in code because zero-price routes, aliases, context limits, and feature support can rotate while an older documentation page remains online. ### Match the quota to the workload shape Translate the allowance into peak requests per minute, input and output tokens, concurrency, retries, and every daily or monthly ceiling that applies to the intended account. The first limit reached by the workload is the practical ceiling. A large token pool can still fail interactive bursts, and a high request limit can still fail long-context work. Include tool calls and retry traffic, test the largest realistic payload, and treat research, experimental, and community access as conditional on their eligibility and fair-use terms. ### Preserve billing and privacy boundaries The recorded payment-card requirement is Unknown, and account access is required. Confirm both in the live signup flow, set provider-side budgets when charges are possible, and observe the balance or usage fields after a complete request. The prompt-handling classification is **Private** because For qualifying GWDG-hosted routes, prompts and responses are not persistently stored or used for AI training; metadata-only logging and narrow operational exceptions remain. Re-read the terms for the exact route and plan before sending sensitive, regulated, or proprietary content. ### Follow the evidence, then re-check it This record links 6 first-party sources covering the offer, catalog, limits, pricing, terms, privacy, or adoption evidence available to the audit. Prefer the newest and most specific governing document or live catalog when sources disagree. A dated finding can become stale even when the page remains online, so retain the source and snapshot that supported the decision and submit a correction when a provider changes a material term. ### Plan fallback without policy drift A fallback should preserve modality, context length, streaming, structured output, tools, safety controls, and data terms, not only API syntax. Decide which errors may retry, cap retry storms, and prevent an exhausted free route from silently switching to a billable model. Rotating aliases and free pools can change behavior without changing the endpoint, so retain the response model field and test at least one substitute route before the primary offer becomes unavailable. ### Monitor the offer as a dependency Capture the model ID, response model, rate-limit headers, usage fields, latency, HTTP status, and provider request identifier. Watch authorization failure, quota exhaustion, catalog removal, policy revision, and balance movement as separate failure modes. Re-check the live catalog and governing sources on a schedule proportionate to the workload's importance, and keep an owner and exit path for any production dependency on volatile free capacity. ### Test one complete request before scaling Start with the smallest permitted request using the exact credential, model ID, endpoint, region, and account type intended for deployment. Record the status, headers, usage fields, response model, latency, and dashboard balance movement. Then exercise an invalid model, quota exhaustion, or rate limit so failure is explicit and cannot silently switch to a paid route. Validate streaming, structured output, and tool calls separately because a free model can expose fewer features than its paid counterpart. Keep a budget ceiling outside the application whenever billing is possible, and do not send sensitive data until the observed route matches the reviewed agreement. ## Primary sources - [KISSKI LLM service](https://kisski.gwdg.de/en/leistungen/2-02-llm-service/): Official Product - [SAIA](https://docs.hpc.gwdg.de/services/saia/index.html): Official Docs - [Available models](https://docs.hpc.gwdg.de/services/ai-services/chat-ai/models/index.html): Official Catalog - [Data privacy notice](https://docs.hpc.gwdg.de/services/ai-services/chat-ai/data-privacy/index.html): Official Privacy - [Chat AI FAQ](https://docs.hpc.gwdg.de/services/ai-services/chat-ai/faq/index.html): Official Docs - [GWDG News 06-2024](https://gwdg.science/en/about-us/gwdg-news/2024/GN_06-2024_www.pdf): Official Report ## Sitemap See the full [semantic sitemap](/sitemap.md) for every page and markdown mirror.