Skip to content

Changelog

v0.4.0-alpha (2026-06-18)

Changed

  • Updated the TULIP serving runtime to Dynamo/vLLM 1.2.1.
  • Migrated Dynamo model deployment manifests to nvidia.com/v1beta1.
  • Refreshed the backend models behind the existing public aliases:
  • chat → google/gemma-4-31B-it-qat-w4a16-ct
  • code → Qwen/Qwen3.6-27B-FP8

Notes

  • API paths and public aliases remain unchanged.
  • The code model currently uses a conservative runtime profile for stability.

v0.3.0-alpha Usability & Access (2026-03-20)

API & Access

  • Introduced model aliases (e.g. chat, code, embed) as the primary user interface
    → Raw model names are no longer user-facing
  • Basic API key authentication enforced at ingress level

Documentation & Examples

  • Complete documentation restructure with:
  • Clear Getting Started flow
  • Capability-based API reference
  • Explicit model mapping
  • Removed outdated or unverified content
  • Added validated, copy-paste examples for core capabilities:
  • Chat, Responses API, Embeddings, Tool calling

Known limitations

  • Tool calling via /v1/responses currently not reliable in this environment due to backend serving limitations
  • Authentication is API-key based only (no SRAM/OIDC yet)
  • No rate limiting or quota enforcement
  • Service remains in pilot phase (no uptime guarantees)

v0.2.0-alpha: API consolidation (2026-03-12)

Added

  • Capability-based API namespaces (/chat/v1, /code/v1, /embed/v1)

API changes

  • Removed embedding endpoint https://embedding.tulip.tudelft.nl
  • Embeddings now available at /embed/v1 under the main API domain
  • All models are now served under the unified domain https://api.tulip.tudelft.nl

Improvements

  • Separated DynamoGraphDeployments per model backend
  • Resolved tool-calling instability caused by mixed model graphs
  • Verified reliable tool-calling for both:
  • openai/gpt-oss-20b
  • Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8

Migration

Old endpoint: https://embedding.tulip.tudelft.nl/v1/embeddings

New endpoint: https://api.tulip.tudelft.nl/embed/v1/embeddings

v0.1.1-alpha: Embedding Support (2026-03-06)

Added

Documentation

  • Updated Getting Started with embedding usage example
  • Clarified separate base URLs for inference and embeddings

Notes

  • Authentication remains disabled during the pilot phase
  • Model availability may evolve as the platform matures

v0.1.0-alpha: Pilot Launch (2025-11-27)

Added

  • OpenAI-compatible endpoints: /v1/chat/completions, /v1/completions, /v1/responses
  • Multi-model serving (Ministral 8B, Qwen2.5 Coder 14B, GPT-OSS 20B)
  • Basic observability: Prometheus, Grafana, DCGM GPU metrics
  • Reference tests repository with curl + Python examples
  • Basic documentation: Landing page + Getting Started + Roadmap
  • GitOps deployment via FluxCD
  • Staging and production environments
  • PVC-backed cache for model storage

Known limitations

  • No authentication yet
  • Streaming unstable for some models
  • Lack of quotas / accounting
  • Limited model portfolio
  • No embedding or multimodal models yet