What is LFM2.5-ColBERT-350M?
LFM2.5-ColBERT-350M is a late-interaction retrieval model from Liquid AI, released on 18 June 2026 with a dense sibling, LFM2.5-Embedding-350M. It keeps a 128-number vector for each token of a document and scores queries by matching tokens (MaxSim), which can be used to search an index or to rerank a shortlist from another system without one. It replaces the deprecated LFM2-ColBERT-350M.
It has about 353 million parameters and is built on LFM2.5-350M-Base, with Liquid's causal model turned bidirectional: attention sees the whole input and the short convolutions look both ways. It covers 11 languages, English, Spanish, German, French, Italian, Portuguese, Arabic, Swedish, Norwegian, Japanese and Korean, and is built for short text: documents up to 512 tokens and queries up to 32.
The weights are on Hugging Face, with GGUF builds for llama.cpp, under the LFM Open License v1.0, which allows free commercial use for organisations below $10 million in annual revenue. It runs through PyLate and Sentence Transformers.
Is LFM2.5-ColBERT-350M worth it?
Cross-lingual short-text search is where this model leads. On Liquid's NanoBEIR multilingual set it averages 0.605 NDCG@10 across 11 languages, against 0.556 for Qwen3-Embedding-0.6B, 0.540 for the older LFM2-ColBERT and 0.484 for LightOn's English-only LateOn. On MKQA, where the query and the answer can be in different languages, it reaches 0.694 recall at 20. English is the exception: LateOn scores 0.690 there to this model's 0.687.
Liquid notes that NanoBEIR scores run about 15% above full BEIR, so these numbers are not comparable with BEIR figures elsewhere, and all of them are Liquid's own runs. LightOn's later mLateOn reports 87.69 on the long-document MLDR benchmark against 78.39 for this model, which fits its design: with a 512-token document cap it is built for product listings, FAQ answers and support snippets, not reports.
Speed is the point of the LFM architecture. On a MacBook Pro with an M4 Max under llama.cpp, encoding a query and scoring it against cached documents takes 8.2 ms at the median; encoding the document as well, as a reranker does, takes 34.3 ms. Liquid's own GPU stack brings that to 22.8 ms with documents encoded and 2.5 ms with them cached. Those latencies make on-device search over files and notes practical without a server.
The licence is the part to read before shipping. The LFM Open License is free for commercial use only below $10 million in annual revenue; larger companies need a separate agreement, which rules it out for many enterprise teams that can take Apache-licensed LateOn or mLateOn instead.
Where it earns its place
A developer at a 30-person e-commerce company selling across Europe and Japan
Find products when a shopper searches in a different language from the listing
Index short product titles and descriptions with LFM2.5-ColBERT-350M in PyLate, which handles cross-lingual queries at 0.694 recall at 20 on Liquid's MKQA test. Below $10 million revenue the licence is free.
An app developer adding local search to a notes app
Search a user's notes and emails on the device, with no server
Run the GGUF build under llama.cpp and keep document vectors on the device; Liquid measures 8.2 ms per query against cached documents on an M4 Max laptop.
An engineer whose agent picks from hundreds of tools
Shortlist the right tools for a request before the LLM sees them
Rerank tool descriptions against the request with PyLate's rank function, as Liquid's tool-selection demo does; descriptions fit well inside the 512-token document limit.
Before you commit
- Setup
- pip install pylate (or sentence-transformers 6.0 or later) and load the model with trust_remote_code=True, which applies the bidirectional patches; GGUF files run in llama.cpp.
- Integrations
- PyLate with FastPLAID indexes, Sentence Transformers' MultiVectorEncoder and llama.cpp; Liquid offers its own GPU serving stack to enterprise customers on request.
- Data and privacy
- Self-hosted, so documents stay on your hardware or the user's device. No hosted API for this model is listed.
- Cost at scale
- No API fee. Organisations with $10 million or more in annual revenue need a commercial licence from Liquid AI, priced on request; storage grows with one 128-dimension vector per token.
- Limits
- Documents are cut at 512 tokens and queries at 32. Eleven languages; Liquid recommends it for short-context retrieval rather than long documents.
What’s good
- 0.605 on NanoBEIR across 11 languages, ahead of Qwen3-Embedding-0.6B
- Strong cross-lingual recall: 0.694 at 20 on MKQA
- GGUF builds; 8.2 ms query scoring on a laptop
- Searches an index or reranks a shortlist with one model
What’s not
- 512-token documents and 32-token queries
- Free commercial use only below $10 million revenue
- Behind mLateOn on long documents and LateOn on English
- Benchmarks are Liquid's own
Best for
Developers building multilingual search over short texts, especially on laptops, phones or other devices without a GPU server.
LFM2.5-ColBERT-350M alternatives
LFM2.5-ColBERT-350M: common questions
- Is LFM2.5-ColBERT-350M worth it?
- The multilingual ColBERT to reach for when passages are short and the search has to run on a laptop or phone, though the licence and the 512-token cap set its limits. We rate it 3.9 out of 5.
- How much does LFM2.5-ColBERT-350M cost?
- Free weights under the LFM Open License v1.0; commercial use is free below $10 million annual revenue, above that a licence from Liquid AI is needed.
- What are the downsides of LFM2.5-ColBERT-350M?
- 512-token documents and 32-token queries; free commercial use only below $10 million revenue; behind mLateOn on long documents and LateOn on English; benchmarks are Liquid's own.
- Who is LFM2.5-ColBERT-350M best for?
- Developers building multilingual search over short texts, especially on laptops, phones or other devices without a GPU server.
- How much does LFM2.5-ColBERT-350M cost for a team or in production?
- No API fee. Organisations with $10 million or more in annual revenue need a commercial licence from Liquid AI, priced on request; storage grows with one 128-dimension vector per token.
- Is LFM2.5-ColBERT-350M safe for company data?
- Self-hosted, so documents stay on your hardware or the user's device. No hosted API for this model is listed.
- What are the best LFM2.5-ColBERT-350M alternatives?
- The closest alternatives we have reviewed are LateOn and mLateOn, Qwen3-Embedding and Qwen3-Reranker, Jina Embeddings v5 and EmbeddingGemma 2.
Answers checked Nov 20, 2026.