GTE-ModernColBERT-v1
LightOn's 150M ModernBERT late-interaction retriever for long documents, trained on fully open data.
About
GTE-ModernColBERT-v1 is LightOn's first-generation ColBERT-style retriever, built on the 149M-parameter ModernBERT backbone and trained with the open-source PyLate library. It stores 128-dimensional token-level vectors and scores with MaxSim late interaction, and although it trained on 300-token documents it generalizes to inputs up to 8,192 tokens.
LightOn reports strong numbers for its size: 83.59 on TREC-COVID and 54.89 on FiQA2018 from BEIR, plus a LongEmbed mean of 88.39, and emphasizes that both pretraining and fine-tuning used fully open data with all signals released as metadata. The weights are Apache 2.0 and the model has over 72,000 Hugging Face downloads. It has since been succeeded by LateOn, but v1 remains the reference open-data late-interaction baseline.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- BEIR: 83.59 TREC-COVID, 54.89 FiQA2018 (author-reported via press coverage)
- LongEmbed mean: 88.39; LEMB Narrative QA: 78.82 (author-reported)
- Successor LateOn reports 57.22 BEIR avg (14 datasets), leading all ColBERT models (author-reported)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: HF API confirms Apache 2.0 license, April 2025 release, and 72K+ downloads; open-data training claims documented on the model card. Strengths: permissive license, genuine adoption, reproducible PyLate recipe, strong small-model results. Docked for: (1) benchmark figures reach me via press coverage of the authors' numbers, not an independent eval; (2) superseded by LateOn, so maintenance attention has moved on.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · HF API confirms Apache 2.0, April 2025 release, 72K+ downloads
- · Fully open-data training claimed on the model card with signals released as metadata
- · Benchmark figures (BEIR 83.59 TREC-COVID, LongEmbed 88.39) reach via press coverage of author numbers; superseded by LateOn