ronaldmannak/LFM2.5-Embedding-350M-8bit

LFM2.5-Embedding-350M — MLX (8-bit)

MLX build of LiquidAI/LFM2.5-Embedding-350M, a multilingual dense bi-encoder (1024-dim CLS embedding, cosine similarity), for local inference on Apple Silicon with MLX.

All weights, architecture, and behavior are LiquidAI's. This repository changes the file format (PyTorch/safetensors → MLX) and post-training quantized to 8-bit (affine, group size 64) from the bf16 MLX conversion. Every Linear and embedding layer is quantized; the non-quantized layers (conv, norms) stay bf16. See the original model card for training details and intended use.

Quantization details

• Quantized with mlx.nn.quantize(mode='affine', bits=8, group_size=64) — the exact configuration benchmarked below.
• Verified bit-exact: the reloaded checkpoint's encodings are identical (max abs diff 0) to the in-memory-quantized model used for the benchmark (verify_export.py) — the shipped artifact is the model measured below.
• Reload applies the quant from config.json["quantization"] before loading weights (see retrieval.load_model).

Evaluation

Retrieval quality of this checkpoint (and its sibling precisions), measured as NDCG@10 / Recall@10 on judged pools. Retention = metric ÷ bf16 metric, averaged per-dataset.

Setup. English = the four NanoBEIR sets (full small corpora, ~2–5k passages, 50 queries each). Multilingual = MIRACL dev (the real queries and relevance judgments) for Spanish, German, Japanese, Arabic, each scored over a reduced pool of ~6k passages (judged positives + hard-mined negatives + sampled distractors, from mteb/MIRACLRetrievalHardNegatives), 100 queries each. Reduced pools make absolute scores easier than full-corpus MIRACL and not leaderboard-comparable — but every precision searches the identical pool, so the retention numbers (the point of this table) are sound. ColBERT uses brute-force MaxSim with no query augmentation, so its absolute scores sit a touch below a full PLAID setup.

Summary (mean over 8 datasets)

precision	NDCG@10	NDCG retention	Recall@10	Recall retention	size
bf16	0.728	100.0%	0.775	100.0%	709 MB
8-bit ◄	0.729	100.1%	0.775	100.0%	377 MB
4-bit	0.730	100.0%	0.766	98.6%	200 MB
mxfp4	0.725	99.8%	0.764	98.4%	—

NDCG@10 by dataset

dataset	bf16	8-bit ◄	4-bit	mxfp4
NanoNQ · en	0.704	0.704	0.703	0.703
NanoFiQA2018 · en	0.504	0.511	0.502	0.498
NanoSciFact · en	0.716	0.717	0.714	0.712
NanoNFCorpus · en	0.342	0.340	0.335	0.345
MIRACL · es	0.891	0.892	0.895	0.893
MIRACL · de	0.809	0.810	0.819	0.812
MIRACL · ja	0.929	0.928	0.940	0.922
MIRACL · ar	0.926	0.926	0.928	0.916

License & attribution

Redistributed under the LFM Open License v1.0 (LICENSE) — the same license as the original model. Per Section 4, this notice records that the files were modified (format conversion to MLX + 8-bit quantization). The original work is by Liquid AI; this repository is an independent conversion, not affiliated with or endorsed by Liquid AI. The license includes a commercial-use threshold (Section 5) — review it for your use case.

Base model: LiquidAI/LFM2.5-Embedding-350M