Skip to content
← Models
Open weightsMcGill-NLP / ServiceNow Research/10 trustMetrics-checked

LLM2Vec

The three-step recipe that turns any decoder-only LLM into a text encoder, no labels required.

Visit official page →

About

LLM2Vec is less a model than a conversion kit. Published in April 2024 by researchers at McGill/Mila and ServiceNow Research, it transforms any pretrained decoder-only LLM into a text encoder in three unsupervised steps: enabling bidirectional attention, masked next-token prediction, and SimCSE-style contrastive learning. The team applied it to models from 1.3B to 8B parameters.

The paper claims a new unsupervised state of the art on MTEB, ranking 6th overall and 1st among models trained only on public data as of May 2024, with supervised fine-tuning pushing it further. The MIT-licensed code ships as a pip package with converted checkpoints on Hugging Face, and the method's influence shows up in later decoder-embedding work like GritLM and the LLM2Vec-Gen follow-up released in 2026.

Specs

Context window

4,096 tokens (base-model dependent)

Pricing

Free / open weights and MIT-licensed code

Free quota

Free / open weights — self-host converted models at no license cost

API access

No official API — pip library (llm2vec) + Hugging Face weights (McGill-NLP collection)

Weights

Open weights

Released

Not disclosed

Benchmarks

Unsupervised MTEB SOTA at release; 6th overall and 1st among public-data-only models (author-reported, qualified claims in the paper)

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

llm-to-encoder-conversionretrieval-researchcustom-domain-embeddingsembedding-fine-tuning

Trust score — 7/10

/10 trustMetrics-checked

Scored 7/10 (metrics tier, metrics-v1). Evidence: arXiv paper (2404.05961, Apr 2024) with the triple-qualified MTEB claims stated in the abstract itself; MIT license and pip package confirmed on GitHub. Strengths: reproducible recipe, permissive license, widely cited method. Docked for: (1) it is a method, not a maintained model artifact, with all figures author-run; (2) later approaches have surpassed its scores; (3) no official hosted access.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
6.0/10
Maintenance · recency, releases, issue load
5.0/10
Openness · license + docs
10.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · arXiv paper 2404.05961 (Apr 2024); McGill/Mila + ServiceNow Research
  • · Triple-qualified MTEB claims (6th overall, 1st among public-data-only) stated in the paper itself
  • · A method/recipe, not a maintained model artifact; later approaches surpassed its scores

Sources