Skip to demo

A brain-like model,
lightweight for the edge.

AwareLiquid is a bio-inspired cognitive AI framework. The attention-free O-Series carries a fixed-size state instead of a growing KV-cache — 0.381 MB at any context length, running CPU-only. Built for lightweight edge deployment, not for the data center.

Carried state0.381 MB
Context1M tokens
vs KV-cache8063× smaller
Get HyperCode Try the live demo
Product

HyperCode — an AI that finishes the work.

Give it one clear instruction and it works for hours until everything is done. 229 built-in domain skills across finance, law, academia and private equity. No VPN, no setup — one-click install on Windows and macOS.

Download HyperCode v0.1.8 · Free trial
Live Demo

Try Both Models in Your Browser

Two distinct surfaces. Science Q&A grounds answers in a 1,000+ item sourced knowledge base (papers with DOIs and citation counts) — it answers only what the sources support, and says so honestly when they don't. M1-128M is the native 128M hybrid trained from scratch with no base model. Switch between them below. Ask in English or Chinese.

Connecting…

Connecting…

Research-grade demo — quality is experimental. What to expect →

Open Full-Page Demo →

Where it fits

Built for streams that never stop, on hardware that cannot grow.

Four deployments with something real behind them — measured, shipped, deployed, or all three. Every number links to open code; anything without data says so.

Battery management

State-of-health regression on voltage/current/temperature streams, on the pack controller itself.

Measured — NASA PCoE cells +7.7% at 80% sample loss vs +31% LSTM/GRU · 2.6 KB streaming state. 10 seeds, whole cell held out.

Hearables

Always-on wake-word spotting where the microphone never closes — one keyword, measured in English. Open source as O1-Sound.

Measured — single keyword only English “hello”: FRR 0.14 ± 0.05 at FAR 0.05 across 3 seeds. 5.03 MB fp32 / 1.27 MB int8, 5,120 B constant state. Not production grade: seed-to-seed spread is wider than any change we have tried, and the multilingual version measured worse than never firing — one binary class cannot span greetings that share no phonetics.

Financial documents

QA over filings far longer than the context window, with the base model frozen behind an API.

Measured — M2 adapter 44/48 (91.7%) across five financial-document domains · ~2.8k tokens per question. A regression baseline, not a leaderboard.

Edge battery safety — in production

Battery AI live at a commercial platform: the liquid core runs on the pack controller itself, alerting offline with no network round-trip.

Deployed — customer production 353-param on-device LTC · 1.3 ms inference · 48 B constant state · offline alerts. Customer-reported: 229 e-buses, 92% early-failure detection, 120.8-day warning lead time.

Data cleaning

Scanned archives, handwritten tables and blurry reports turned into clean structured data — a commercial service on the same liquid core.

Measured — service pilot 100% table-digit accuracy, zero hallucination · ~1h per 189-page report. See the before/after cases on the service page.

Also targeting In-vehicle systems · On-device assistants · Industrial robotics — architectural fit (bounded state over an unbounded session, native Δt), no pilot data yet.

Model Lineup

Every Release, One Place

Each AwareLiquid release states its status plainly: live (served on this site, try it now), download (weights published, run it yourself), or training (checkpoint exists, run not finished). No version is silently half-shipped.

M-Series ● Live on site

M1-128M

From-scratch hybrid — attention + liquid core in every layer. 50K-step run on WikiText-103. English text continuation.

128.6M params12 layers · 832d512-token window
Try it live → Download weights ↓ Served by the native decoder on this site
M-Series ● Live on site

M1 · TinyLlama-1.1B + MT v2s

Frozen TinyLlama-1.1B-Chat base with trained liquid-core residual adapters + LoRA. The conversational model — English and Chinese, with fast-weight memory across turns.

1.1B base + 2.3M adapterchat-capablebilingual EN/ZH
Try it live → Download adapter ↓ 24.8 MB checkpoint · replicate the live demo
O-Series ↓ Download

O1-48M

Attention-free O-series edge model, trained from scratch. Constant 0.381 MB carried state regardless of context — the O(1) memory line. Run it on CPU.

48M paramsattention-freeCPU-friendly
Download weights ↓ GitHub release · serve with server.py
M-Series ◐ Training

M1-2B

1.9B-parameter hybrid (2912d × 35L) training on A100. A learning-rate scheduling bug that stalled convergence is fixed — validation PPL is now falling steadily (107 → 74.75, -30%). Weights will be published when the run converges.

1.9B params35 layers · 2912dconverging · PPL -30%
Weights when it converges Training on rented A100
Research Family

Clone & Reproduce

The wider family beyond the served models — each repository ships real code and a measured test suite. License: MIT.

O-Series ↓ Clone & run

O1-Anti — The Anti-Transformer

Three pillars against Transformer costs: Neural Liquid Adjacency (8× smaller cache), sparse module graph (27% activation), non-autoregressive decode (3.4× fewer passes). MQAR recall (wd=0.1): ties attention at 1.000; Mamba 0.07-0.10. 22/22 tests.

efficiency line22 testspip install torch
GitHub →
O-Series ↓ Clone & run

AwareLiquid-Physic

Physics written into the architecture: liquid (LTC) substrate + hard-constraint Hamiltonian head + symplectic integration. Project complete — all PRD targets met: semigroup training lifts the liquid advantage to 66% (mean of 7 configs), few-shot adaptation 3.24×, N-body / non-separable / probabilistic / time-conditioned heads. 43/43 tests.

physics-native43 testsMIT
GitHub →
Cortex ↓ Clone & run

C1 — Five-Layer Cognition Stack

Sensory stream → O(1) memory → sparse temporal routing → predictive coding → liquid GRPO + System 1/2. E1 liquid GRPO first run (supervised arm): extrapolation 1.0→0.56 (seed0). E1–E3 auto-bootstrap dependencies from the public M1 mirror.

pure-liquid visionE1–E3 experimentsMIT
GitHub →
O-Series ↓ Clone & run

O1-Sound — Wake-Word Spotter

Always-on keyword spotting on the liquid core: 5,120-byte constant state, ONNX export 5.03 MB fp32 / 1.27 MB int8. English validated (FRR 0.146 @ FAR 0.046). 32/32 tests.

edge · always-on32 testsONNX 5MB
GitHub → Checkpoint ↓
Research ↓ Read

MT-LNN Technical Paper

English + 中文. Microtubule-inspired liquid network: O(1) working memory, multi-scale predictive coding, compute skipping. Empirical tables are canonical — honest verdicts included.

EN / 中文PDFopen science
PDF (EN) → PDF (中文) →
Research ↓ Explore

LLM-Visualization

Interactive 3D visualization of the MT-LNN architecture (extended from bbycroft/llm-viz).

3D explorerwebMIT
GitHub →