AI Potluck
Back to Gap Map Model components / Embeddings & retrieval models

Nomic Embed

Nomic AI
open source / Overall score: 3.6

Nomic Embed produces long-context text embeddings for search and retrieval-augmented generation. The v1 and v1.5 encoders accept up to 8,192 tokens, and v1.5 is trained with Matryoshka representations so a vector can be truncated to 64 dimensions with graded loss of quality. Version 2 is a mixture-of-experts encoder covering roughly 100 languages at a 512-token limit. Nomic AI publishes weights, training data and the contrastors training pipeline together.

Scored over the nomic-embed-text line only. The first-party GGUF conversions are the same checkpoints in another container and are not counted twice, and nomic-embed-code, nomic-embed-vision and modernbert-embed-base are separate releases. Verified 2026-09-11 via the Hugging Face model cards, the Hub API and the contrastors README.

Openness

5 high confidence
5.0
weights
open(Apache-2.0 ungated checkpoints for v1, v1.5 and v2-moe)
data
open(the card states the training data is released in its entirety, served through the contrastors repository behind a free Atlas account)
code
open(contrastors ships MLM pretraining, contrastive pretraining and finetuning entry points with their configs, plus the scripts that generate the data)
license
Apache-2.0(OSI)

The full-release case this rung was written for: weights, corpus and the pipeline that consumed it, all under Apache-2.0, with the v2 paper stating the point of the release is reproducibility of the training run. The corpus is fetched from a Nomic bucket after registering for an Atlas account, which is friction at the download gate rather than a corpus withheld.

Adoption

5 high confidence
5.0

16,124,079 downloads in the trailing 30 days for nomic-embed-text-v1.5 alone, plus 3,623,688 for v1 and 2,193,696 for v2-moe. The Hugging Face universe sweep puts the three shipped nomic-embed-text checkpoints at 22,689,628 combined, excluding the first-party GGUF conversions of the same weights.

Capability

3 high confidence
3.0

Solid and widely used, and measurably behind the frontier encoders on the same leaderboard table - 62.28 against 64.38 for a model of similar size. The v2 mixture of experts claims parity only within its parameter class, and gives up the long-context window the v1 line was known for.

Verified 2026-09-11