LFM2.5 Encoder 350M Code MXFP8-GPTQ

This is a modified RESMP.DEV research release derived from LiquidAI/LFM2.5-Encoder-350M at revision b886781f7c6f10ca9b7096e21b83e30a073c2f39. It is not an official Liquid AI release. We removed the masked-language-model head and contrastively fine-tuned the full bidirectional encoder for multilingual code retrieval.

Quantization finding

This is a research artifact, not an automatic recommendation to replace the BF16 model. Activation calibration is compared with matched native round-to-nearest quantization and the complete machine-readable receipts are included so mobile and Apple-Silicon users can evaluate the size, latency, memory, and quality tradeoff themselves.

Held-out retrieval results

All rows use the same untouched 6,995-pair multilingual test set, 1,200-character query and 4,000-character passage caps, query token cap 512, and passage token cap 2,048. Higher is better. RTN is a matched quantization control; Nomic and Jina are external service baselines, not architecture-matched controls.

A separate BF16 cross-runtime run on NVIDIA GeForce RTX 3090 Ti with PyTorch 2.13.0+cu130 produced MRR 0.3709, 616.2 queries/s, 139.1 passages/s, and 1109.0 MB peak CUDA allocation. CUDA throughput is reported separately and is not compared directly with Metal.

A paired 10,000-sample bootstrap estimates calibrated MXFP8 minus BF16 MRR at +0.0005, with a 95% interval of [-0.0010, +0.0021]. A point estimate whose interval crosses zero is not presented as a quality win.

Usage

git clone https://github.com/RESMP-DEV/calibrated-code-embeddings
cd calibrated-code-embeddings
uv sync --extra mlx
CODE_EMBEDDING_MODEL_PATH=/path/to/this-model code-embedding-serve --port 1235

The service exposes POST /v1/embeddings. It runs the bidirectional LFM2.5 body directly with MLX; LM Studio is not required. Prefix retrieval queries with query: and candidate code with passage: when calling the model directly.

Training and data receipts

Full-backbone symmetric in-batch InfoNCE training used 24,626 language-balanced pairs selected from the 42,626-row source training split, two epochs, batch size 32, learning rate 2e-5, temperature 0.05, and seed 17. The training report records the NVIDIA RTX A6000 runtime and validation history.

• train: 42,626 rows, SHA-256 426ebfaad34b14d7627ba6e668ae36e08e548c9d057b0edc208bcfa6fe527629
• validation: 5,319 rows, SHA-256 9ac88b3138de4ca94c2ef3a87ccf19381fc76265c2bc9d65b4791983d0315096
• test: 6,995 rows, SHA-256 9ed10842a12132b6bfb5421df1e2f88dbcfbf6f6e960f36b22eb9ea6e3c72315
• calibration: 4,096 rows, SHA-256 ee9edaf80a6854c18053b96521090a51bdb76642abeb98618d7aed36e70b6de9

The corpus combines pinned CodeSearchNet data with pinned permissively licensed code repositories. Exact and token 8-gram near-duplicates were removed with test-before- validation-before-train precedence. See corpus_receipt.json, source_receipt.json, training_report.json, quantization_report.json when present, benchmarks/, and artifact_manifest.json for machine-readable evidence.

License and attribution

The weights retain the LFM Open License v1.0 in LICENSE, including its attribution and commercial-use conditions. MODIFICATIONS.md identifies RESMP.DEV's changes. The training and quantization workbench is separately MIT licensed.

Citation

@article{liquidAI2026Encoders,
  author = {Liquid AI},
  title = {LFM2.5-Encoders: Fast at Long Context, Even on CPU},
  journal = {Liquid AI Blog},
  year = {2026},
  note = {www.liquid.ai/blog/lfm2-5-encoders},
}