Oct 19 7 days ago 23:18 Ship int8 safetensors with per-row scales for FastEmbed Stores the released int8 rows as and their per-row scales as , so FastEmbed's static path reproduces the released encoder exactly. Restores to the released version; FastEmbed now handles special tokens and sqrt pooling itself. 16:24 fix: make it usable by fastembed 16:18 chore: conversion and test script, make special tokens in ZeroEncoder configurable 16:14 feat: add static-embeddings like F16 safetensors lookup table Oct 7 19 days ago 17:41 update fastembed installation instructions Sep 25 21:58 Update Hub IDs to the Qdrant org Sep 16 13:15 docs: remove repository demo link 13:12 docs: link runnable Constella demo 12:57 docs: keep benchmark reporting focused on established results 12:55 docs: note full BEIR benchmark in progress 12:50 docs: explain the Constella family and recommended retrieval setup 12:44 docs: simplify and align model card Sep 15 22:29 Update model card for the research preview: one FastEmbed branch for the whole family, native fp32 query vectors, canonical clean-4 results with all six beside them. Sep 4 20:28 constella-zero — FastEmbed-first card, run p35w-2m-s2500 16:35 Update README.md 04:33 constella-zero — FastEmbed-first card, run p35w-2m-s2500 04:21 constella-zero — FastEmbed-first card, run p35w-2m-s2500 03:19 constella-zero — FastEmbed-first card, run p35w-2m-s2500 Sep 3 23:29 zero v1 — M7 lookup table, run p35w-2m-s2500 23:28 zero v1 — M7 lookup table, run p35w-2m-s2500 23:19 zero v1 — M7 lookup table, run p35w-2m-s2500 23:02 zero v1 — M7 lookup table, run p35w-2m-s2500 15:37 zero v1 — M7 lookup table, run p35w-2m-s2500 15:23 zero v1 — M7 lookup table, run p35w-2m-s2500 15:23 initial commit 
