# Multiverse Computing — CompactifAI API

> Multiverse Computing, S.L. (Donostia-San Sebastian, Spain) operates the CompactifAI API,
> an OpenAI-compatible LLM inference service serving tensor-network-compressed versions of
> frontier open models alongside uncompressed third-party models. Base URL
> https://api.compactif.ai, with EU (https://api-eu.compactif.ai) and US
> (https://api-us.compactif.ai) data-residency endpoints. Bearer API key auth, one key per
> account, pay-as-you-go per-token billing.

Generated: 2026-08-26
Method: generated
Source: apis.yml plus the artifacts in this repository. No llms.txt is served by the
provider — /llms.txt returns 404 on api.compactif.ai and on multiversecomputing.com, and
403 (S3 AccessDenied) on docs.compactif.ai.

## API surface

13 operations, OpenAPI 3.1.0, published unauthenticated at
https://api.compactif.ai/openapi.json

- POST /v1/chat/completions — chat completion (streaming via `stream: true`, SSE)
- POST /v1/completions — text completion
- POST /v1/responses — OpenAI Responses API; `store: true` persists, owner-scoped
- GET  /v1/responses/{response_id} — retrieve a stored response
- POST /v1/audio/transcriptions — Whisper speech-to-text (multipart, streaming supported)
- POST /v1/files — upload a file (multipart); no delete operation exists
- GET  /v1/files/{file_id}/content — download raw file bytes
- POST /v1/batches — create an async batch job from an uploaded input file
- GET  /v1/batches — list batch jobs (cursor pagination: `after`, `limit` 1–100)
- GET  /v1/batches/{batch_id} — retrieve a batch and its live status
- POST /v1/batches/{batch_id}/cancel — cancel an in-progress batch
- GET  /v1/models — list available model IDs
- GET  /v1/models/{model} — model information

## Models

Compressed by CompactifAI: cai-mistral-small-3-1-slim, hypernova-60b, carina-60b,
qwen-3-6-27b, cai-whisper-large-v3-turbo-slim.
Third-party, uncompressed: gpt-oss-120b, mistral-small-3-1, nemotron-3-nano-omni, glm-5-1,
cai-glm-5-1, glm-5-2, quasar-438b.
Catalog: https://docs.compactif.ai/models/

## Authentication

`Authorization: Bearer <API key>`. One key per account, issued from
https://dashboard.compactif.ai/ after billing setup, or via AWS Marketplace subscription
through Multiverse IAM (manually approved, typically within 24 hours). Rotate from Manage
API Keys. No OAuth, no scopes. https://docs.compactif.ai/authentication/

## What an agent needs to know before calling

- No idempotency. No Idempotency-Key header exists, and ChatCompletionRequest declares no
  `seed`. A retried write is executed and billed again. Reconcile with GET /v1/batches or
  GET /v1/responses/{response_id} rather than blind-retrying.
- No rate limits are published, no RateLimit-*/Retry-After headers are returned, and 429 is
  not documented or declared in the spec. There is no runtime signal to back off on.
- One reversal path exists (cancel a batch) and no window is stated for it. Every inference
  call is irreversible and metered on receipt.
- No dry-run, no cost estimate, no tokenizer endpoint. You cannot price a call before
  making it.
- Errors use two envelopes: `{"error":{type,message,param,code}}` on documented errors, and
  a bare `{"detail": ...}` on 401, 404, 422 and 500. `param` and `code` are documented as
  always null.
- Every response carries `x-request-id`; you may also supply it on the request. Also
  returned: `x-served-by-region`, `x-envoy-upstream-service-time`.
- Model IDs are withdrawn without notice — eleven have been removed since launch, one of
  them ten weeks after being added. Call GET /v1/models rather than hard-coding.
- Zero Data Retention: prompts and completions are not stored; anonymized request logs are
  deleted after 30 days. https://docs.compactif.ai/data-privacy/

## Documentation

- Developer portal: https://docs.compactif.ai/
- Introduction: https://docs.compactif.ai/introduction/
- Quickstart: https://docs.compactif.ai/quickstart/
- API reference: https://docs.compactif.ai/api_reference/
- Authentication: https://docs.compactif.ai/authentication/
- OpenAI compatibility: https://docs.compactif.ai/openai-compatibility/
- Error handling: https://docs.compactif.ai/error-handling/
- Data privacy: https://docs.compactif.ai/data-privacy/
- FAQ: https://docs.compactif.ai/faq/
- Changelog: https://docs.compactif.ai/changelog/
- Pricing: https://docs.compactif.ai/pricing/
- Support: https://docs.compactif.ai/support-contact/
- Status: https://status.compactif.ai/
- Community (Discord): https://discord.com/invite/KfqDt3En

## Integrations

Configured as an OpenAI-compatible base URL, not as first-party SDKs: Claude Code, Cursor,
GitHub Copilot, OpenCode, LiteLLM, Hermes Agent, OpenClaw.
Multiverse Computing publishes no first-party client library on npm, PyPI or any other
registry, and its GitHub organization (github.com/multiverse-computing) has 0 public repos.

## Not available

No MCP server. No A2A agent card. No GraphQL. No webhooks or AsyncAPI. No gRPC/Protobuf.
No SOAP/WSDL. No /.well-known document on any host. No Postman collection. No sandbox or
test-mode keys. No first-party CLI. No published SLA. No public roadmap.

## Company

- Website: https://multiversecomputing.com/
- Product: https://multiversecomputing.com/compactifai
- Blog / resources: https://multiversecomputing.com/resources
- Legal notice (terms of use): https://multiversecomputing.com/legal-notice
- Privacy policy: https://multiversecomputing.com/privacy-policy
- AWS Marketplace: https://aws.amazon.com/marketplace/pp/prodview-ce2dp5ibd7lli
- Other product: Singularity (quantum and quantum-inspired optimization) — no public
  developer documentation.