Hugging Face's logo

buggandhi
/
Orqanta-AI

Text Generation
English
edge-ai
automotive
function-calling
electric-vehicles
two-wheeler
embedded
cpu-only
Model card Files
xet
 Community

Orqanta AI

Orqanta AI (Psycho Engine)

Ultra-Lightweight 121M Foundation Model for Electric & Petrol Vehicles

License: Apache-2.0 Hugging Face Parameters Model Size RAM Footprint Latency GPU Requirement Target Mobility Python

⚡ Executive Summary

Orqanta AI (powered by the native C++ Psycho Engine) is an ultra-fast, 100% offline edge AI foundation model purpose-built for automotive intelligence, in-vehicle infotainment (IVI), and digital instrument clusters.

Engineered with 121 Million Parameters and packed into a 33.7 MB zero-overhead .cact binary container, Orqanta AI runs entirely on embedded CPUs, requiring less than 80 MB of RAM with sub-50ms inference times (guaranteed execution under 2 seconds on any automotive ECU or cluster SoC with $\le$ 2GB RAM).

Key Highlights

  • 🧠 121M Parameter Architecture: 20-layer Grouped-Query Attention (GQA) transformer equipped with 18,432-slot associative Engram memory.
  • 🏎️ Deterministic Sub-50ms Latency: Generates structured function calls in 15ms – 40ms on commodity ARM64 / x86_64 CPUs.
  • 💾 Microscopic RAM Footprint: Consumes only ~78.9 MB peak RAM, leaving 90%+ system memory free for cluster graphics, safety RTOS, and GPS navigation.
  • 🚫 100% Offline & Air-Gapped: Zero cloud dependency, zero cellular latency, and zero telemetry leakage — fully operational in remote highways, underground tunnels, and dead zones.
  • 🛡️ Byte-Level Grammar Decoding: Mathematically guarantees 100% syntactically valid JSON function calls conforming strictly to predefined vehicle tool schemas.
  • 🔌 Zero External GPU/NPU Needed: Runs efficiently on pure CPU cores without specialized accelerators.

🚗 Supported Vehicle Domains

Orqanta AI provides out-of-the-box support for four primary mobility segments:

Vehicle Type Propulsion Target Hardware Supported Telemetry & Controls
Cars Electric (EV) Android Automotive OS (AAOS), Linux IVI Battery thermal telemetry (BAT-015), SOC degradation, charging curves, regen braking, cabin HVAC
Cars Petrol / ICE / Hybrid Head Units, OBD-II Gateway Fuel flow rate, engine RPM, coolant temp, DTC fault codes (P0420, P0128, U0100), service intervals
Two-Wheelers Electric Scooters ARM64 Digital TFT Cluster Battery SOC %, real-time range estimation by riding mode (Eco / Ride / Sport), motor temp, TPMS alerts
Two-Wheelers Petrol Motorcycles Smart Digital Cluster / Speedo Fuel tank level, Distance-To-Empty (DTE), real-time mileage (km/L), trip telemetry, chain slack warnings

🔬 Exact Model Parameters & Specifications

The Orqanta base model (models/orqanta.cact) is configured with the following audited parameters:

================================================================================
  Orqanta AI (Psycho Engine) — 121M Architecture Specification
================================================================================
  • Total Parameters:          121,021,910 (~121.02 Million / 121M)
  • Quantized Weights (CQ):    120,394,752 (~120.39M) [Codebook Quantized 4-bit]
  • High-Precision Weights:    625,110 (~0.63M) [FP16 Norms, Gates & Embeddings]
  • Transformer Layers:        20 Layers
  • Hidden Dimension (d_model):768
  • Attention Heads:           12 Query Heads (qk_dim = 48, v_dim = 64)
  • Key-Value Heads:           2 Heads (Grouped-Query Attention for low bandwidth)
  • Context Window:            8,192 Tokens
  • Vocabulary Size:           8,192 Tokens (Automotive & Telemetry BPE)
  • Engram Memory Slots:       18,432 Associative Gathering Slots
  • On-Disk Binary Size:       33.70 MB (Direct mmap, zero parsing overhead)
  • Active Peak RAM:           78.9 MB
================================================================================

📊 Performance Benchmarks

Metric Orqanta AI (121M) TinyLLaMA / Qwen (1B – 1.5B) Cloud APIs (GPT-4o / Claude)
Model Size 33.7 MB 1.8 GB – 3.2 GB Multi-Gigabyte (Cloud)
Peak Inference RAM ~78.9 MB 1.6 GB – 2.8 GB N/A (Requires Internet)
Target Hardware RAM $\le$ 1GB – 2GB RAM Minimum 4GB – 8GB RAM Internet Modem Required
Hardware Required Commodity CPU Only NPU / High-end GPU recommended Cloud Datacenter GPU
First Token Latency 15 ms – 40 ms 1,200 ms – 3,500 ms 600 ms – 2,500 ms
Vehicle SLA Compliance 100% (< 2.0s SLA guaranteed) Variable (Prone to timeouts) Fails in signal dead zones
Decode Speed 95 – 120 tok/sec (CPU) 18 – 35 tok/sec (CPU) Streamed over HTTP
Tool Calling Guarantee 100% Valid Grammar JSON Frequent schema hallucinations High latency retries

🏛️ Model Architecture 

[ Natural Language Voice / Telemetry Input ]
                    │
                    ▼
┌────────────────────────────────────────────────────────┐
│             SentencePiece BPE Tokenizer                │
│             (8,192 Vocabulary Partition)               │
└────────────────────────────────────────────────────────┘
                    │
                    ▼
┌────────────────────────────────────────────────────────┐
│     20-Layer Laddered Attention Network (d=768)        │
│                                                        │
│  ├── Grouped-Query Attention (12 Q-Heads / 2 KV-Heads) │
│  ├── Monarch Hadamard Fast Feed-Forward Blocks        │
│  └── 18,432-Slot Associative Engram Memory Layers      │
└────────────────────────────────────────────────────────┘
                    │
                    ▼
┌────────────────────────────────────────────────────────┐
│       Byte-Level Grammar Constrained Decoder           │
│   (Forced JSON Schema Output Matching Tool Rules)      │
└────────────────────────────────────────────────────────┘
                    │
                    ▼
[ Deterministic Vehicle Property & CAN-bus Action Frames ]
  1. Monarch Hadamard Factorization: Replaces dense projection matrices with block-diagonal butterfly structures, retaining rich parameter representation while minimizing floating-point compute.
  2. Grouped Query Attention (GQA): Pairs 12 query heads with 2 KV heads, dramatically reducing the KV-cache footprint during autoregressive token generation.
  3. Engram Associative Memory: Incorporates n-gram associative memory tables that perform fast memory gathers, enabling the compact 121M parameter model to achieve accuracy on par with 500M–1B models.
  4. Byte-Level Constrained Decoding: Applies grammar masks directly to output logits, ensuring that generated tokens strictly follow valid JSON syntax and match vehicle function signatures.

📦 Repository Structure 

Orqanta-AI/
├── bin/
│   ├── linux-x86_64/orqanta     # Native x86_64 C++ binary (Linux IVI & Dev testing)
│   └── android-arm64/orqanta    # Native ARM64 C++ binary (AAOS & Instrument Clusters)
├── models/
│   └── orqanta.cact             # Audited base model binary (33.7 MB, 121M parameters)
├── Orqanta_AI_Dataset_v1/        # 360-record SFT Car Telemetry Dataset (EV & Petrol)
│   ├── train.jsonl              # 288 training chat dialogues
│   ├── validation.jsonl         # 36 validation dialogues
│   ├── test.jsonl               # 36 held-out test dialogues
│   ├── orqanta_ai_dataset.csv   # Flat CSV representation
│   └── DATASET_CARD.md          # Dataset schema & safety specifications
├── Orqanta_TwoWheeler_Dataset_v1/# 380-record SFT Two-Wheeler Dataset (Bikes & Scooters)
│   ├── train.jsonl              # 304 training dialogues
│   ├── validation.jsonl         # 35 validation dialogues
│   ├── test.jsonl               # 41 held-out test dialogues
│   ├── two_wheeler_dataset.csv  # Flat CSV representation
│   └── DATASET_CARD.md          # Maintenance, battery & TPMS specifications
├── include/
│   └── orqanta.h                # C/C++ Header for embedded firmware integration
├── lib/
│   ├── linux-x86_64/liborqanta.a# Static C++ engine library for Linux x86_64
│   └── android-arm64/liborqanta.a# Static C++ engine library for Android ARM64
├── orqanta_ai/                  # Python SDK & fine-tuning runtime
│   ├── model/                   # Architecture, fine-tuning & export pipelines
│   │   ├── finetune.py          # Standalone LoRA fine-tuning CLI
│   │   ├── architecture.py      # JAX/Flax 121M Transformer definition
│   │   └── export.py            # .cact deployment packager
│   └── runner.py                # Subprocess & binary bridge
├── examples/
│   ├── car_android/             # In-car HVAC, GPS navigation & EV/ICE telemetry
│   │   ├── orqanta_car_assistant.py # Offline car telemetry & DTC diagnostic assistant
│   │   ├── car_assistant.py     # Real-time climate, windows, and navigation controller
│   │   ├── car_tools.json       # Automotive vehicle property schemas
│   │   └── deploy_car_adb.sh    # ADB push script for Android Automotive
│   ├── two_wheeler/             # Scooter & Motorcycle cluster assistant
│   │   ├── two_wheeler_assistant.py # EV battery range & petrol mileage cluster assistant
│   │   └── two_wheeler_tools.json # Two-wheeler instrument cluster schemas
│   └── voice_assistant.py       # Vehicle cockpit mic & helmet intercom voice pipeline
└── scripts/
    └── eval_dataset.py          # Dual dataset benchmark & latency validator

⚡ 1-Command Download & Run

No compilation, no pip dependencies, and zero cloud API keys needed. You can clone and run Orqanta AI on your Linux PC, laptop, or embedded ECU in a single copy-paste command:

🚗 Car Assistant: DTC Diagnostics & AC Control (1-Command) 

git clone --depth 1 https://github.com/GANDHIKALASI/Orqanta-AI.git && cd Orqanta-AI && ./bin/linux-x86_64/orqanta --model models/orqanta.cact --tools examples/car_android/car_tools.json --prompt "Check diagnostic trouble codes and set temperature to 21 degrees"

🛵 Two-Wheeler Cluster: EV Battery & Range (1-Command) 

git clone --depth 1 https://github.com/GANDHIKALASI/Orqanta-AI.git && cd Orqanta-AI && ./bin/linux-x86_64/orqanta --model models/orqanta.cact --tools examples/two_wheeler/two_wheeler_tools.json --prompt "Check my electric scooter battery percentage and estimated range in Eco mode"

📱 Android Automotive / ARM64 Cluster (1-Command ADB Deploy) 

adb push bin/android-arm64/orqanta models/orqanta.cact examples/car_android/car_tools.json /data/local/tmp/ && adb shell "chmod +x /data/local/tmp/orqanta && /data/local/tmp/orqanta --model /data/local/tmp/orqanta.cact --tools /data/local/tmp/car_tools.json --prompt 'Check high voltage battery temperature and set AC to 20'"

🤗 Download Base Model from Hugging Face Hub

The official 121M base weights are published on Hugging Face: buggandhi/Orqanta-AI.

1. Download in 1 line using Hugging Face CLI: 

hf download buggandhi/Orqanta-AI models/orqanta.cact --local-dir .

2. Or download directly in Python: 

from huggingface_hub import hf_hub_download

# Downloads models/orqanta.cact (33.7 MB) directly from Hugging Face
model_path = hf_hub_download(repo_id="buggandhi/Orqanta-AI", filename="models/orqanta.cact")

🚀 Quickstart & Usage

1. Standalone Binary CLI (Zero Dependencies)

Run directly from your terminal:

./bin/linux-x86_64/orqanta \
  --model models/orqanta.cact \
  --tools examples/car_android/car_tools.json \
  --prompt "Check diagnostic trouble codes and set temperature to 21 degrees"

Real Response (< 40ms, 79MB RAM):

{
  "type": "call",
  "success": true,
  "function_calls": [
    {"name": "query_diagnostic_trouble_codes", "arguments": {"subsystem": "all"}},
    {"name": "set_climate", "arguments": {"temperature": 21}}
  ],
  "reasoning": "DTC diagnostics -> query_diagnostic_trouble_codes; temp 21 -> set_climate 21.",
  "confidence": 0.942,
  "peak_ram_mb": 78.9,
  "prefill_tps": 95.4,
  "decode_tps": 99.8
}

2. Python SDK

Install in development mode:

pip install -e .

Define vehicle tools with type hints and run:

import orqanta

@orqanta.tool
def adjust_cabin_climate(temperature: int, fan_speed: int = 3):
    """Adjust in-car climate temperature in Celsius and blower fan speed."""
    print(f"Applying HVAC target: {temperature}°C (Fan: {fan_speed})")
    return {"status": "applied", "target_temp": temperature}

# Initialize agent (automatically links local binary & models/orqanta.cact)
agent = orqanta.Orqanta(tools=[adjust_cabin_climate])

# Execute command
result = agent.run("Set the car AC to 20 degrees with fan speed 4")
print(result["results"])
# [{'status': 'applied', 'target_temp': 20}]

3. Interactive Demos

🚗 In-Car Diagnostic & Climate Assistant 

python3 examples/car_android/orqanta_car_assistant.py

🛵 Two-Wheeler Cluster (EV Range & Petrol Mileage) 

python3 examples/two_wheeler/two_wheeler_assistant.py

🎙️ Cockpit & Bluetooth Helmet Intercom Voice Pipeline 

python3 examples/voice_assistant.py

📊 Dual SFT Vehicle Datasets

Orqanta AI includes 740 curated vehicle dialogue records partitioned for training, validation, and testing:

python3 scripts/eval_dataset.py

====================================================================
  🚗 Car (EV & Petrol) SFT Dataset Overview (Orqanta_AI_Dataset_v1)
====================================================================
  • Total Examples:    360 (Train: 288 | Val: 36 | Test: 36)
  • Coverage:          DTC diagnostics (P0420, P0128, U0100), battery thermal 
                       alarms, energy curves, ADAS event filtering, CAN/HAL.

====================================================================
  🛵 Two-Wheeler SFT Dataset Overview (Orqanta_TwoWheeler_Dataset_v1)
====================================================================
  • Total Examples:    380 (Train: 304 | Val: 35 | Test: 41)
  • Coverage:          EV scooter battery SOC & range by mode (Eco/Ride/Sport),
                       petrol motorcycle mileage (km/L), TPMS tyre pressure.
====================================================================

Fine-Tuning CLI

To train custom vehicle adapters on your proprietary fleet telemetry:

python3 -m orqanta_ai.model.finetune Orqanta_AI_Dataset_v1/train.jsonl \
  --epochs 3 \
  --batch-size 16 \
  --lr 1e-4 \
  --lora-rank 16

🔧 C / C++ Embedded Integration

Embed Orqanta AI directly into vehicle firmware and ECU daemons using include/orqanta.h and lib/liborqanta.a:

#include "orqanta.h"
#include <stdio.h>

int main() {
    char out_buffer[4096];
    int rc = orqanta_complete("Check battery temperature and range", NULL, 0, 512, out_buffer, sizeof(out_buffer));
    if (rc >= 0) {
        printf("Vehicle Action: %s\n", out_buffer);
    }
    return 0;
}

Compile with zero external dependencies:

gcc -O3 main.c -Iinclude -Llib/linux-x86_64 -lorqanta -lm -lpthread -o vehicle_agent

👨‍💻 Developer & Maintainer

Gandhi Kalasi

Gandhi Kalasi

AI Product Engineer Aspirant | Creator & Maintainer of Orqanta AI

Passionate about building AI-powered products, exploring large language models (LLMs) and AI agents, and developing practical software solutions that turn complex ideas into useful, real-world applications.

GitHub Profile   Instagram   Hugging Face   Repository   Issues

📄 License

This project is licensed under the Apache-2.0 License - see the LICENSE file for details.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
Text Generation
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support