LFM2.5-ColBERT-350M

Name: LFM2.5-ColBERT-350M
Author: LiquidAI

by LiquidAI

Edge-sized ColBERT late-interaction retriever built on LFM2.5

3Kdl/month

117likes

353Mparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

LiquidAI/LFM2.5-ColBERT-350M

Feature URI

mixpeek://reranker@v1/lfm25_colbert_350m_v1

Overview

LFM2.5-ColBERT-350M is Liquid AI's compact late-interaction retriever: instead of collapsing a document into one vector, it keeps a small vector per token and scores relevance with MaxSim between query and document token vectors. That preserves term-level precision that single-vector embeddings blur away, which shows up on exact-phrase, entity-heavy, and long-document retrieval. At 350M parameters with the LFM2.5 backbone's efficiency, it runs late-interaction quality at edge and on-CPU budgets where classic ColBERT deployments were impractical.

On Mixpeek, a late-interaction model like this slots into the reranking or precision stage of a multi-stage retriever: a dense first stage recalls candidates cheaply, then token-level MaxSim re-scores the top-K. See the late interaction retrieval guide for when token-level matching beats single vectors and what it costs at the index layer.

Architecture

ColBERT-style late-interaction architecture (PyLate-compatible) on the LFM2.5 hybrid backbone: per-token contextual embeddings with MaxSim scoring, trained for sentence-similarity and retrieval. 350M parameters, English-focused, runs via sentence-transformers/PyLate with custom code enabled.

Mixpeek SDK Integration

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so LFM2.5-ColBERT-350M runs
// on your side and the output is upserted through POST
// /v1/namespaces/{namespace_id}/documents/upsert. On Enterprise the other
// path is to upload the weights instead: POST /v1/namespaces/{id}/models
// accepts the huggingface format and a custom plugin loads them.
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection_id: "col_your_collection",
      documents: [
        {
          document_id: "asset-00412",
          // The vector name has to match a vector index on the collection.
          vectors: { "text-embedding": yourVector },
          payload: { source_key: "archive/2026/asset-00412" },
        },
      ],
    }),
  },
);

// Managed alternative, if this exact model is not the requirement:
// text_extractor@v1 runs intfloat/multilingual-e5-large-instruct
// (1024-d) over a bucket, with no inference of your own.

Capabilities

Token-level late-interaction scoring (MaxSim) for precise term matching
Multi-vector document representations that survive exact-phrase queries
Edge/CPU-friendly footprint at 350M parameters
PyLate and sentence-transformers compatible for drop-in retrieval stacks
Strong fit as a precision reranking stage over a dense first stage

Use Cases on Mixpeek

Reranking top-K candidates from a dense retriever with token-level precision

Entity- and phrase-heavy corpora (legal, technical docs) where single vectors blur terms

On-prem or edge retrieval where large rerankers do not fit

Hybrid stacks pairing BM25 or dense recall with late-interaction re-scoring

Benchmarks

Dataset	Metric	Score	Source
Hugging Face	Pipeline tag	sentence-similarity (ColBERT/PyLate)	Liquid AI, 2026 Model Card
Hugging Face	Parameters	350M	Liquid AI, 2026 Model Card
Hugging Face	Backbone	LFM2.5	Liquid AI, 2026 Model Card