cstr/lfm2-colbert-GGUF

LFM2.5-ColBERT-350M — CrispEmbed GGUF

CrispEmbed-native GGUF quantizations of LiquidAI/LFM2.5-ColBERT-350M.

Multi-vector (ColBERT-style) retrieval: per-token embeddings projected to 128 dimensions, L2-normalized. Uses late interaction (MaxSim) scoring for fine-grained token-level matching.

Format note: These GGUFs use CrispEmbed's internal tensor naming (lfm.* prefix, arch=lfm2). They include the colbert.projection.weight tensor from the 1_Dense module. Not compatible with llama.cpp.

Model variants

Architecture

• —Backbone: LFM2.5-350M bidirectional hybrid (16 layers: 10 ShortConv + 6 GQA attention, 1024-dim hidden, SwiGLU FFN)
• —ColBERT head: Linear(1024, 128) + L2 normalize per token
• —Scoring: MaxSim — max over doc tokens of cosine similarity per query token, summed
• —Parameters: 350M + 128K projection head
• —Languages: EN, ES, DE, FR, IT, PT, AR, SV, NO, JA, KO (11 languages)
• —Task prefixes: "query: " for queries, "document: " for passages

Usage

# ColBERT multi-vector encode
./crispembed -m lfm2-colbert-q8_0.gguf --colbert "query: what is deep learning?"

# JSON output (per-token vectors)
./crispembed -m lfm2-colbert-q8_0.gguf --colbert --json "query: machine learning"

# Server
./crispembed-server --embed lfm2-colbert-q8_0.gguf --port 8080
curl -X POST http://localhost:8080/colbert/score \
  -d '{"query": "what is deep learning?", "documents": ["Deep learning is a subset of ML", "The weather is nice"]}'

from crispembed import CrispVit

model = CrispVit("lfm2-colbert-q8_0.gguf")
assert model.has_colbert

# Encode multi-vector representations
query_vecs = model.encode_multivec("query: what is deep learning?")   # (n_tokens, 128)
doc_vecs = model.encode_multivec("document: Deep learning uses neural networks")

# MaxSim scoring
score = model.maxsim(query_vecs, doc_vecs)
print(f"Score: {score:.4f}")

use crispembed::CrispEmbed;

let mut model = CrispEmbed::new("lfm2-colbert-q8_0.gguf", 4)?;
assert!(model.has_colbert());

let query = model.encode_multivec("query: what is deep learning?");
let doc = model.encode_multivec("document: Neural networks learn representations");

Conversion

Convert from the source model yourself:

git clone https://github.com/CrispStrobe/CrispEmbed
cd CrispEmbed

# Convert (loads 1_Dense/model.safetensors for ColBERT projection)
python models/convert-lfm2-embed-to-gguf.py \
    --model LiquidAI/LFM2.5-ColBERT-350M \
    --output lfm2-colbert-f32.gguf --dtype f32

# Quantize
./build/crispembed-quantize lfm2-colbert-f32.gguf lfm2-colbert-q8_0.gguf q8_0
./build/crispembed-quantize lfm2-colbert-f32.gguf lfm2-colbert-q5_k.gguf q5_k
./build/crispembed-quantize lfm2-colbert-f32.gguf lfm2-colbert-q4_k.gguf q4_k

License

LFM Open License v1.0 — same as the base model.

Credits

Original model by LiquidAI. GGUF conversion and inference engine by CrispEmbed.

Provenance and EU AI Act Art. 53 note

• —Upstream model: LiquidAI/LFM2.5-ColBERT-350M — published by LiquidAI.
• —Upstream licence: other. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• —What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• —Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
• —Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.