Models and regions
isms-mini, isms-fast, and isms-thinking aliases, including EU-oriented -eu variants.
Inhalt auf Englisch
The API exposes a small set of model aliases. You pass the alias as the model field on /v1/chat/completions. Behind the alias, ISMS Copilot routes to a compliance-tuned model stack and injects framework knowledge when relevant.
Aliases
| Alias | Intent |
|---|---|
isms-mini | GLM-5.3-Flash (global). Low-cost bulk lane: formatting, extraction, classification, and other high-volume work that does not need deep analysis |
isms-fast | GLM 5.3 (global). Fast answers, lower latency |
isms-thinking | GLM 5.3 (global). Deeper reasoning for harder questions |
isms-fast-eu | Mistral (EU). Fast path with EU-oriented routing |
isms-thinking-eu | Mistral (EU). Thinking path with EU-oriented routing |
The four standard aliases (isms-fast, isms-thinking, and their -eu variants) share the same unit price ($/M). Choosing a region does not change the list rate. Thinking usually costs more in practice because it uses more tokens (including reasoning tokens that are not shown), not because the list rate is higher. isms-mini is a cheaper bulk model on the global path. Live unit rates are shown in the platform Models and Pricing pages.
isms-mini is built for volume, not depth: evidence formatting, field extraction, classification, and JSON normalization are its lane. It has no thinking mode. For analysis, drafting, and anything a reviewer signs, use the fast or thinking aliases.
Exact upstream providers can change. The aliases are the stable contract for your integration. Why these models, and the evaluation record behind them: Model quality and grounding.
Thinking and max_tokens
Thinking aliases (isms-thinking, isms-thinking-eu) spend part of the completion budget on internal reasoning before the visible answer. If you set a very low max_tokens (for example under a few hundred), the budget can be used up before any answer text is produced.
For those thinking aliases, the API raises a client max_tokens below 1024 up to 1024 so the request can still return content. Fast aliases (isms-fast, isms-mini) still spend a small reasoning budget, so a client max_tokens below 256 is raised to 256. If you omit max_tokens, the API uses its normal high default (tens of thousands of tokens).
When a raise happens, the response includes:
| Header | Meaning |
|---|---|
x-isms-max-tokens-requested | The value you sent |
x-isms-max-tokens-effective | The value actually used (at least 1024 on thinking, 256 on Fast/mini) |
You may be billed for output up to the effective budget, not only the lower number you requested.
Regions and dual-mode
- Prefer
-eualiases when you need EU-oriented processing for a request. - Bare
isms-fastandisms-thinkinguse global processing when global mode is on (it is on in production). isms-miniis always global. It is not an EU alias. Callers who need EU-only processing should use-eu, not mini.
For the latest operator-facing wording about providers and data residency, see the Trust Center, Zero Data Retention, and the in-console docs at platform.ismscopilot.com/docs.
Framework knowledge
Completions are compliance-aware: the stack injects framework reference knowledge from named-framework detection on user messages and the last assistant turn, or exactly the modules you pin. You control it per request with ismscopilot: { "frameworks": "auto" | "none" | ["ISO_27001", ...] }, enumerate coverage with GET /v1/frameworks, and every response discloses what was injected via the x-isms-frameworks header and the ismscopilot response object. The server prompt that frames it is published word for word: see The API system prompt, in full. Why this is not any generic model API, plus honest detection limits: Framework knowledge. You do not maintain a separate RAG store of standards text for the built-in frameworks.