Zero-Cloud Sovereign Binary Registry

Compressed Models & LoRA Cartridges

Explore our pre-quantized 1.58-bit and 2-bit foundation models and hot-swappable ~22 MB LoRA cartridges. All models load in 0 ms via zero-copy memory mapping (`mmap`).

Compressed Foundation Models

Android App Tested & Verified

Qwen 2.5 1.5B (2-Bit INT8-Embed)

Our battle-tested tripartite foundation model with INT8 embeddings (890 MB). Formatted in pure-Rust .leviathan2 binary for instantaneous zero-copy mmap. The exact, certified model powering our Android Chat App on Snapdragon 6 Gen 1 (18–21 TPS sustained).

Target ApplicationLeviathan Android Chat App
Physical RAM Footprint890 MB (Sub-1GB Mobile)
Sustained Decode Speed20.3 TPS (Snapdragon 6 Gen 1)
Time-to-First-Token< 20 ms (Warm Multi-Turn)
Zero-Throttling Temp32.8°C Certified Continuous
Binary Download Size890 MB (.leviathan2 format)
Desktop / Laptop / High-RAM Mobile • NEW

LLaMA 3.1 8B (1.58-Bit Ternary)

Meta's flagship LLaMA 3.1 8B compressed from 16 GB (FP16) to just 2.05 GB via our proprietary 1.58-bit ternary quantization pipeline — an 87.2% reduction. Runs entirely on consumer hardware without a GPU. Recovers near-original accuracy through SVD residual correction and God Mode LoRA adaptation.

Original Model Size16.0 GB (FP16 Baseline)
Compressed Binary Size2.05 GB (.leviathan2 format)
Compression Ratio87.2% Smaller (1.58-bit Ternary)
Mobile Speed (Sustained)11.6 TPS (Moto G Stylus 5G Tested)
Desktop Speed84.5 TPS Sustained (Apple M3 Max)
Quality RecoverySVD Residual + God Mode LoRA V2
Target HardwareDesktop, Laptop, 8GB+ Mobile
GSM8K (Math Reasoning)74.8% (with God Mode LoRA)

Hot-Swappable "God Mode" LoRA Cartridges (~22 MB)

Swap intelligence domains on the fly in 0 ms without restarting the inference engine. Each cartridge modifies attention and MLP projections directly in memory.

FEATURED
Flagship Universal Adapter

God Mode LoRA V2 (Universal Reasoning)

Trained across 3 full passes of SFT + DPO alignment over 12 elite datasets. Elevates our tested Qwen 2.5 1.5B model to competitive 7B-class GSM8K math and HumanEval coding benchmarks with 0 ms load overhead.

Binary Size:22.4 MB
Capacity:Rank 256 / Alpha 512
Target:Qwen 2.5 1.5B (2-Bit)
Load in SDK →
Software Engineering

Autonomous Code & Algorithmic Logic

Specialized in Python, Rust, TypeScript, and C++ code generation, complex refactoring, AST parsing, and zero-hallucination API compliance.

Binary Size:21.8 MB
Capacity:Rank 256 / Alpha 512
Target:Qwen 2.5 1.5B (2-Bit)
Healthcare & Diagnostics

Clinical Emergency Triage Cartridge

Offline clinical emergency diagnostic heuristics, Glasgow Coma Scale, vital sign threshold matrices, and pharmaceutical contraindication guardrails.

Binary Size:18.5 MB
Capacity:Rank 128 / Alpha 256
Target:Qwen 2.5 1.5B (2-Bit)
Automotive & Telemetry

Automotive OBD-II & Field Diagnostics

Full SAE J1979 diagnostic trouble code graph (DTC P0000-P0999), sensor normal operating bands, and automated failure-tree diagnosis for field mechanics.

Binary Size:14.2 MB
Capacity:Rank 128 / Alpha 256
Target:Qwen 2.5 1.5B (2-Bit)
Gaming (UE5 / Unity)

Gaming NPC & Dynamic Dialogue (120 FPS)

Zero-heap allocation conversational intelligence certified for 120 FPS game-loop execution. Multi-faction memory and quest dialogue trees.

Binary Size:16.0 MB
Capacity:Rank 128 / Alpha 256
Target:Qwen 2.5 1.5B (2-Bit)
Legal & Regulatory

Legal Clause Analysis & Compliance

Air-gapped NDA, master service agreement, and regulatory clause extraction with high-precision compliance verification.

Binary Size:20.1 MB
Capacity:Rank 256 / Alpha 512
Target:Qwen 2.5 1.5B (2-Bit)