cufa64/lln_v16_blend_100k

LLN v16 — Living Language Network (3-Corpus Blend)

Zero-parameter language generation from pure graph topology.

No neural networks. No gradient descent. No learned weights.

Model Details

Property	Value
Vocabulary	100,000 tokens
Forward edges	117.5M directed
PMI edges	34.5M
Trigram pairs	11.9M
Total bigrams	6.42 billion
Corpus	FineWeb-Edu (10GB) + Gutenberg (12GB) + OpenWebText (10GB)
Build time	~3 hours (CPU)
Model size	2.1 GB
Format	LMDB (data.mdb)

Usage

pip install numpy lmdb huggingface_hub
git clone https://github.com/andycufari/lln
cd lln
python generate.py --prompt "The fire burned"
# The model auto-downloads on first run

Examples

Prompt	Output
The king	a more powerful voice heard my lord hath commanded thee thy life . The most famous letter addressed a
She opened the door	opened fire on the door swung open . She paused abruptly closed the " " She laughed softly closed doors
The fire burned	alive , and the public safety training camp with his eyes flashed brightly glowing cheeks
The ship sailed	northward along the right to go straight to him so forth a big leagues farther inland navigation channel the whole
Scientists discovered	they would not believe that he began studying all I hope you don't expect to think you find you want

Architecture

Generation uses a 5-phase pipeline:

1. Frequency-penalized PMI activation — identifies semantic targets, suppresses rare-word hallucination
2. Flow-aware target selection — avoids topological dead ends using local push/receive mass
3. Beam search grammar walk — 5 competing paths find bridges across low-weight gaps
4. Target depletion — reached targets are zeroed, activation landscape shifts
5. Natural halting — generation stops when semantic energy is exhausted

See the whitepaper for full details.

License

MIT — Andy Cufari, 2026