LFM2.5 Retrievers
Open-source multilingual retrieval models built for on-device search
LFM2.5 Retrievers are bidirectional semantic search models released by Liquid AI on June 18, 2026, designed for fast multilingual retrieval across 11 languages. Both models—ColBERT and Embedding variants—run on CPU laptops, phones, and edge servers, eliminating the need for expensive APIs or cloud infrastructure.
These compact retrieval models solve a specific problem: organizations need to search knowledge bases, product catalogs, and FAQ collections quickly and affordably while keeping data on-device. Neither model requires a GPU or cloud call, making them ideal for privacy-sensitive workflows, cost-constrained deployments, and teams building retrieval at the edge.
What Is LFM2.5 Retrievers?
LFM2.5 Retrievers are two compact semantic search models built by Liquid AI, released on June 18, 2026. Both hold 350M parameters and represent bidirectional members of the LFM (Liquid Foundation Models) retrieval family. The two variants—LFM2.5-ColBERT-350M and LFM2.5-Embedding-350M—serve different speed-vs-accuracy tradeoffs.
ColBERT uses per-token vectors, matching queries word-by-word for higher accuracy on complex questions. Embedding converts each document to a single vector, prioritizing speed and a smaller index. Both support Arabic, German, English, Spanish, French, Italian, Japanese, Korean, Norwegian, Portuguese, and Swedish.
Weights are available free on Hugging Face under the LFM Open License v1.0, permitting commercial use by companies under $10M revenue. No API subscription or cloud infrastructure is required.
LFM2.5 Retrievers offer cost-free deployment for qualifying companies, but integrating them into your workflow requires architecture decisions around indexing, latency, and compliance. Discover how to build privacy-first RAG pipelines using LFM2.5 Retrievers.
Book a ConsultationKey Features & Capabilities
LFM2.5 Retrievers excel at short-context retrieval: product catalogs, FAQ systems, support documentation, and internal knowledge bases. Both run on CPU with low latency, requiring no GPU or cloud infrastructure. Multilingual search spans 11 languages with explicit cross-lingual training, enabling queries in one language to retrieve documents in another.
GGUF formats ship with llama.cpp, MLX, vLLM, and ONNX support, simplifying on-device deployment. Model size fits efficiently on phones, laptops, and embedded systems. No authentication, API keys, or rate limits apply to self-hosted instances.
Inference is fast on consumer hardware, comparable to or faster than cloud API calls. Batch retrieval on GPU infrastructure achieves additional speedups for high-volume use cases.
Benchmark Performance
Both models performed competitively on standard multilingual retrieval benchmarks. ColBERT achieved higher accuracy than the Embedding variant on tasks involving complex semantic matching. Per-token matching improves accuracy through training techniques including contrastive pretraining, cross-lingual distillation, and fine-tuning.
On multilingual question-answering tasks, both models showed strong cross-lingual transfer, meaning queries in one language reliably retrieve documents in other supported languages. Accuracy scales with training data quality and the breadth of the target corpus.
Inference latency is competitive on consumer CPU hardware compared to cloud-based retrieval APIs. GPU deployment provides additional speedups for batch retrieval scenarios common in enterprise RAG systems.
Pricing & Licensing
LFM2.5 Retrievers are free to download, run, and fine-tune for companies with under $10M in annual revenue. No per-token API costs, licensing fees, or runtime royalties apply during that threshold. Weights are public on Hugging Face under an open license.
For companies exceeding $10M revenue, a commercial license is required; contact Liquid AI's sales team for pricing terms. Nonprofits and 501(c)(3) organizations can use the models for non-commercial work with no revenue cap.
Self-hosting eliminates API costs entirely. One-time hardware cost replaces subscription fees, making operational expense near-zero for organizations with steady retrieval volume.
Where LFM2.5 Retrievers Excel
E-commerce product search: retailers use the Embedding variant to index catalogs and run queries with low latency. Cross-language support means a customer searching in one language finds products available in another, improving conversion and user experience.
Customer support: knowledge base retrieval using ColBERT powers FAQ bots that match support tickets to resolution docs with token-level precision. On-device deployment keeps customer data local, avoiding third-party APIs and reducing latency.
Enterprise knowledge work: teams deploy retrievers inside Slack plugins or browser extensions to search internal docs (legal contracts, design specs, code repositories) without sending queries to cloud providers. Privacy-sensitive organizations prefer on-device inference for compliance and data control.
Limitations & Tradeoffs
These are retrieval-only models; they do not answer questions or generate content. Pairing with a larger LLM (like Claude) is required for full RAG pipelines. Performance depends on document quality and relevance—garbage in, garbage out.
ColBERT's per-token approach requires larger indices than the Embedding variant, adding storage and search overhead. For small document sets, this is negligible; for large-scale enterprise corpora, index size matters. The Embedding variant trades recall accuracy for speed and compactness.
The models support 11 specified languages with explicit training. Other languages receive no formal support; performance degrades gracefully without published benchmarks. Models are optimized for short queries and documents under 1,000 tokens.
How It Compares to Alternatives
Compared to open-weight retrieval models: LFM2.5 is designed specifically for multilingual and edge-device deployment, whereas many alternatives focus on English-only or cloud-optimized retrieval.
Against paid APIs (OpenAI Embeddings, Cohere Embed): LFM2.5 has zero per-token cost and full local control, with no rate limits or dependency on cloud provider uptime. Trade-off: you manage infrastructure and accept responsibility for the retrieval quality.
Against legacy BM25 search: neural retrievers (LFM2.5 or otherwise) capture semantic meaning, retrieving answers to paraphrased queries. BM25 matches exact keywords only. Neural retrievers deliver superior relevance for natural-language queries at comparable or better latency than legacy keyword search.
How to Get Started
Download weights from Hugging Face under the Liquid AI namespace: LiquidAI/LFM2.5-ColBERT-350M or LiquidAI/LFM2.5-Embedding-350M. Use the Sentence-Transformers library for Python integration. No sign-up or API key is required.
For portable deployment, use GGUF versions and run via llama.cpp, MLX, or vLLM. Mobile and embedded inference is supported through these frameworks. Deployment guides and example code are available on Liquid AI's documentation site.
Integrate into your RAG pipeline: embed documents offline, store vectors in a local vector database (FAISS, Qdrant, or SQLite), and retrieve at runtime. Start with a small corpus to verify accuracy before scaling to production.
Business Use & Compliance
LFM2.5 models are available under an open license permitting commercial use and modification for qualifying companies. On-device inference eliminates cloud data residency concerns, keeping sensitive information local.
For regulated industries (finance, healthcare, law), on-device deployment is a compliance advantage. However, model outputs are not warranted for medical diagnosis, legal advice, or high-stakes financial decisions—human review and governance are required for YMYL (Your Money or Life) use cases.
Enterprise support, formal SLAs, and ISO certifications are not available in the open-source tier. For high-stakes deployments, consult with AI specialists or Liquid AI's enterprise team regarding compliance, testing, and integration best practices.
Frequently Asked Questions
- ColBERT converts each token into a vector, enabling word-by-word query matching for higher accuracy on complex questions, at the cost of larger indices. Embedding turns each document into a single vector, offering the fastest search and smallest index at the cost of lower recall. Choose ColBERT for accuracy; Embedding for speed.
- LFM2.5 Retrievers are free to download and run if your company has under $10M in annual revenue. No per-token API fees, subscription costs, or licensing fees apply. Companies above $10M revenue require a paid commercial license from Liquid AI.
- Yes, the LFM Open License v1.0 permits commercial use, modification, and distribution for companies generating under $10M in annual revenue. If your company exceeds that threshold, contact Liquid AI's sales team for a commercial license. Nonprofits can use the models for non-commercial work with no revenue cap.
- LFM2.5 has zero per-token cost and runs locally without cloud infrastructure, giving you full data control and no rate limits. OpenAI and Cohere offer managed reliability and scaling, but charge per million tokens. LFM2.5 is better for cost-sensitive and privacy-critical deployments; APIs suit high-volume, high-reliability scenarios.
- LFM2.5 explicitly supports 11 languages: Arabic, German, English, Spanish, French, Italian, Japanese, Korean, Norwegian, Portuguese, and Swedish. Cross-lingual search allows queries in one language to retrieve documents in others. Other languages may work but have no official support or benchmarks.
- Yes. Both models run on CPU with low latency on standard consumer hardware. GGUF versions work with llama.cpp and MLX for mobile deployment. No GPU or cloud infrastructure is required, making it ideal for offline-first and edge applications.
- On-device inference keeps data local, eliminating cloud residency concerns important for compliance. However, the open-source model does not come with warranties, SLAs, or formal certification. For regulated use (healthcare, finance, law), you must implement additional governance, human review, and testing.
- If your company revenue exceeds $10M USD, the LFM Open License v1.0 no longer grants commercial rights. You must obtain a commercial license from Liquid AI. Pricing for that tier is not publicly available; contact their sales team directly.
Ready to Deploy LFM2.5 Retrievers in Your Workflow?
Adopting LFM2.5 Retrievers requires decisions about indexing, latency targets, and compliance governance. Layer3 Labs helps you architect on-device RAG pipelines, evaluate open-source models, and navigate licensing thresholds for your team.
Book a Consultation