embedme/lightonai-lateon-code-edge-f16

LateOn-Code-edge (GGUF f16 + Projection)

GGUF conversion of lightonai/LateOn-Code-edge for use with litembeddings.

Model Details

Property	Value
Base model	lightonai/LateOn-Code-edge
Architecture	ModernBERT (17M params)
Output dimensions	48 (after projection)
Context length	8,192 tokens
Quantization	f16
GGUF size	34 MB
Projection	256 → 48 (composed from two PyLate Dense layers: 256→512→48)
Use case	Fast, CPU-friendly code search with late interaction (ColBERT-style)

Variants

Variant	Size	Quality
f32	66 MB	Original precision (lossless)
f16 (this repo)	34 MB	Lossless — 100% top-1 agreement, 240/300 weighted
Q8_0	19 MB	79% weighted score, 96-100% top-1 agreement, 3.5× smaller

Files

File	Size	Description
lightonai-lateon-code-edge-f16.gguf	34 MB	ModernBERT encoder in GGUF f16 format
lightonai-lateon-code-edge-f16.projection	49 KB	Composed projection matrix (48×256, float32)

Usage with litembeddings

.load ./build/litembeddings

-- Load model with projection
SELECT lembed_model('lightonai-lateon-code-edge-f16.gguf',
    '{"colbert_projection": "lightonai-lateon-code-edge-f16.projection"}');

-- Generate token embeddings for code
SELECT lembed_tokens('async fn get_connection(pool: &Pool) -> Result<Connection>');

-- Code search with MaxSim
SELECT
    id, code,
    lembed_maxsim(lembed_tokens('database connection pool'), token_emb) AS score
FROM code_embeddings
ORDER BY score DESC
LIMIT 10;

Quantization Quality Benchmark

Tested across 3 codebases (jq/C, Rails/Ruby, FastAPI/Python) with 150 questions total (15 easy + 20 medium + 15 hard per codebase). Weighted scoring: easy×1, medium×2, hard×3 = 100 points per codebase, 300 total.

Aggregate Weighted Scores

Variant	Weighted Score	Percentage
f32	240 / 300	80.0%
f16	240 / 300	80.0%
Q8_0	237 / 300	79.0%

Per-Corpus Scores

Corpus	f32	f16	Q8_0
jq (C)	66/100	66/100	63/100
Rails (Ruby)	79/100	79/100	79/100
FastAPI (Python)	95/100	95/100	95/100

Quantization Quality (Top-1 Agreement vs f32)

Corpus	f16	Q8_0
jq	100.0%	96.0%
Rails	100.0%	100.0%
FastAPI	100.0%	98.0%

Key Findings

• f16 is lossless — identical weighted score (240/300) and 100% top-1 agreement across all codebases
• Q8_0 loses only 1% — 237/300 vs 240/300, drops only on hard queries in jq corpus
• Q8_0 is fastest — 2.5s avg query vs 3.4s f32 vs 13.4s f16 (CPU without FP16 hardware)
• Easy/medium questions show zero quality difference between all variants

Conversion

Converted using litembeddings' ColBERT converter with PyLate projection support:

python scripts/convert_colbert_to_gguf.py lightonai/LateOn-Code-edge ./models \
    --name lightonai-lateon-code-edge-f16 --quantize f16

The converter automatically detects the PyLate two-layer projection structure (1_Dense + 2_Dense) and composes them into a single projection matrix via W_composed = W2 @ W1.