MT-LNN — Microtubule-Inspired Liquid Neural Network (O1-48M)
Honest status first: research artifact, not a general assistant. This is the O1-48M attention-free edge model trained from scratch — the O-Series line of the AwareLiquid research project. It does not aim to match frontier models on dialogue or benchmarks; its claims are about memory form-factor and efficiency, and every number below is measured and reproducible.
What it is
- 48M params, attention-free, from scratch — no base model
- O(1) constant inference state: 0.381 MB flat state from 512 to 1,048,576 tokens (vs a KV cache growing to 3,072 MB — 8,063× smaller)
- Continuous-time liquid core with learnable time constants (13 protofilaments × 5 time-scales), true parallel-scan recurrence
- CPU-friendly: the live demo at awareliquid.ai/demo runs on CPU
Measured results (multi-seed, reproducible)
| Result | Number |
|---|---|
| Cross-window associative recall (fast-weight memory) | 0.56 (3 seeds) vs 0.000 for attention/LoRA |
| O(1) inference state @1M tokens | 0.381 MB constant, 8,063× smaller than KV |
| Robustness to irregular sampling (NASA battery SoH) | +7.7% degradation @80% dropped samples (LSTM +31.1%, GRU +32.8%) |
| Language-modeling quality | Not an advantage: 125M-scale WikiText-103 PPL 88.93 ± 0.33 vs Transformer 78.86 ± 0.25 |
The architecture's case rests on memory form-factor and efficiency, not on quality per parameter. Full honest analysis: RESULTS.md.
Weights
This checkpoint is published on GitHub Releases: o1-48m-v1.
Serve it:
CKPT_PATH=o1_48m_serve.pt TOKENIZER=gpt2 python -m uvicorn serve.server:app
Paper
Related
- Code: AwareLiquid/M1 (MIT)
- Live demo: awareliquid.ai/demo
- Model family: awareliquid.ai