Nemotron-3.5-Embed-1B-mlx-8bit

8-bit integer affine quantization of the text model's linear weights (group size 64), produced with MLX; the vision tower is not used for text embedding. Mean pooling, 2048 dimensions, query: and passage: prefixes required, as the source. Loads through the mlx-vlm fork's text-only mistral3 embedding loader (ivan-avramov/mlx-vlm). Measured on the Premisa retrieval study (English corpus, 230 questions): equal to gte-modernbert-base on every delivered-hit figure at 2000-character chunks and pool 50; multilingual coverage per the source card, not measured by Premisa. Peak task-server memory beside a 1.5B task model and the 8-bit reranker: 3.87 GB.

Source: nvidia/Nemotron-3.5-Embed-1B-BF16 at revision e24ca1e346bcc53572bd7bdc02b3e8ebf66c197c (source weights SHA-256 ad3e1b8bbda04bbd8c969f19df0b66ea672f915f624800ba0b982c5c01bf92af), quantized on 2026-10-22 with Premisa's model_quant_spike.py quantize --bits 8 (mode affine, group size 64). Serving: the mlx-vlm fork's task server (python -m mlx_vlm.server --embedding-model or --reranker-model with this directory); the pooling configuration is read from 1_Pooling/config.json where present.

License: the source model's license applies to these derived weights (openmdw-1.1); see the source card for its terms. Quantization and measurements by Premisa (ai.premisa.desktop); figures are pre-adjudication study measurements, not quality claims.