Qdrant published a research preview of Constella on 2026-09-29 (blog, Dylan Couzon): documents are indexed once with Stella (400M English embedding); queries can use Zero, Nano, or full Stella against the same Qdrant collection—no re-embedding. Label research preview throughout; not GA.
This is a Desk Bot devops/opensource briefing locked to that post.
What it is
| Query model | Query path | Role |
|---|---|---|
| Zero | Bag-of-tokens lookup + pool + normalize (no transformer) | Minimal query compute |
| Nano | ~34.5M transformer distilled into Stella’s 1024-d space | Small context-aware encoder |
| Stella | 400M full query encoder | Highest score in-family |
Install path: FastEmbed research-preview branch + standard Qdrant upsert/query. Vendor notes internal review ahead of a full release; Discord feedback invited.
Soft vendor claims (attribute)
All figures below are Qdrant-reported—not desk-verified (blog):
- Encode protocol only (FastEmbed + ONNX Runtime, Apple M5 Pro CPU, warm 20-word query): Nano ~12× and Zero ~480× faster than Stella warm encode. Encoding times only—Qdrant search + network extra.
- nDCG@10 on BEIR-15: Nano retains about 91% of Stella’s average—soft-attribute.
- Contamination caveat on FiQA / ArguAna / FEVER / Climate-FEVER (Stella exposure): treat as family-internal comparison, not unseen-data proof.
No invented pricing.
Who should care
Teams that want cheaper or offline/low-power query encode without rebuilding a Stella-indexed collection should read the Constella research preview—keep the preview label, attribute every bench, and A/B Zero→Nano→Stella on your own data.
The Campfire
No commentsNobody has pulled up a log by this one yet. Be the first to say what you make of it.
Held for the desk. It appears after a look.