JiRackTernary_1b

💎 Llama 3 JiRack Ternary 1.5B ( Final Release + JiRack Turbo Quant )

High-Performance Ternary-Quantized Transformer on CPU | [PATENT PENDING]

Inventor: Konstantin Vladimirovich Grabko
Organization: CMS Manhattan Official Site: www.cmsmanhattan.com

• Use Pro version https://huggingface.co/CMSManhattan/JiRackTernaryPro_1b

JiRack Ternary 1.5B – Efficient Cloud & Edge Expert Model

JiRackTernary-1.5B is a lightweight, high-performance cloud-optimized model designed to deliver excellent results while significantly reducing cloud inference costs. It serves as an ideal expert model in RAG systems and can be deployed efficiently using the ONNX JiRack Java Server as a powerful alternative to heavier solutions.

Key Features & Resources:

• Full Model Assets: Complete model code, architecture details, training steps, and weights are provided in the repository folder.
• Live Support: Get real-time assistance and chat with the model via email at [email protected].
• JiRack Android Chat Client (with voice input and Ollama API support):
  Download APK
• JiRack Windows 11 Desktop Chat Client (with Ollama API support):
  Download ZIP

JiRack Codding Agent IDE

• It is Agent Coding IDE for JiRack Models to run via Ollama on home PC
• It good choose for Agent Coding IDE such as Cursor , Windsurf IDE or Devin IDE etc but more safe that ask you to apply changes and review.
• Test version https://huggingface.co/CMSManhattan/JiRackDeltaNet_27b/resolve/main/jirack_ide.zip
• Final release version https://huggingface.co/CMSManhattan/JiRackDeltaNet_27b/resolve/main/jirack_ide_final.zip

Recommended Training Approach:

• Best & Easiest Way: Use LoRA for training, then convert the packed model + adapter to ONNX format for optimal deployment and performance.
• Standard SFT Note: Full Supervised Fine-Tuning (SFT) on small models carries a higher risk of overfitting.

Training Process

The model is trained using Quantization-Aware Training (QAT). During training, the weights learn to live in ternary format while being stored in FP16. After training, the pack.py script converts them into real ternary weights.

The resulting model can then be converted to ONNX for production use, or to GGUF for Llama.cpp. However, the best runtime performance is currently achieved with BitNet.cpp, which offers the strongest support for the JiRack architecture.

JiRack Ternary with Default CPU

• Simple run with JiRack UI in docker

docker run -d --name jirack_ternaty_1b -p 7869:7869 --restart unless-stopped cmsmanhattan/jirack_ternaty_1b:latest

---

services:

image: cmsmanhattan/jirack_ternaty_1b:latest
container_name: jirack_onnx_service
ports:
  - "7869:7869"
volumes:
  - .:/app
  - ./web:/app/web
environment:
  - MAX_TOKENS=1024
  - TEMPERATURE=0.7
  - TOP_P=0.9
  - DEFAULT_STREAM=False
  - INTRA_THREADS=4
  - USE_ENV_ALLOCATOR=1
deploy:
  resources:
    limits:
      memory: 16g   

Access the UI

• Once the container is running, open your browser and navigate to:

http://localhost:7869

JiRack Coder 7B is free

• Free coding model from JiRack: https://huggingface.co/CMSManhattan/JiRackCoderQwen_7b

JiRack 10B is available exclusively under a commercial enterprise license.

• RAG model https://huggingface.co/CMSManhattan/JiRack_10b

JiRack Coder 32B is available exclusively under a commercial enterprise license.

• JiRack Coder 32B https://huggingface.co/CMSManhattan/JiRackCoderReasoing_32b

• JiRack Coder 14B https://huggingface.co/CMSManhattan/JiRackCoderReasoing_14b

• JiRack Coder 8B https://huggingface.co/CMSManhattan/JiRackCoderReasoing_8b

---

⚠️ IMPORTANT NOTICE — PROPRIETARY TECHNOLOGY

This model and all accompanying code, algorithms, and documentation are proprietary technology owned by Konstantin Vladimirovich Grabko.

© 2025 Konstantin Vladimirovich Grabko. All Rights Reserved. Patent Pending.

Allowed:

• Personal and non-commercial research use only

Strictly Prohibited without a written commercial license:

• Any commercial use (SaaS, mobile apps, edge devices, paid services, etc.)
• Creating and distributing derivative models for profit
• Removing or modifying any copyright or legal notices
• Patenting any part of this technology

Commercial users must obtain a signed license and pay 5% royalty on net revenue.

Any unauthorized commercial use will be pursued legally under New York law.

Contact for commercial license: [email protected] There is fix price for FinTech

⚠️ Finch tech AL solution

Custom AI Solutions with JiRack

• Deploy your own secure, high-performance model from scratch. I specialize in delivering the JiRack modern architecture on NVIDIA Clusters, fully optimized for your private datasets.
• Let's build your sovereign AI today. DM for inquiries.
• Please contact to CMS Manhttan for the solution

⚠️ Intellectual Property Notice

The architecture, weights, and methods (BRE, SWA Fusion, and HD-FFN) contained herein are the proprietary intellectual property of Konstantin Vladimirovich Grabko.

• Status: Patent Pending (U.S. & International Claims Filed).
• Usage: Commercial use requires a signed execution of the CMS Manhattan JiRack License V.1.2.

---

🚀 Project Overview: The 1.5B "Wide-FFN" Advantage

JiRack-1.5B is a redesigned architecture powered by the JiRack BitNet v2.0 Specification. It is specifically optimized for high-throughput inference on non-NVIDIA hardware (AMD ROCm/HIP) and cloud-native environments (AWS Lightsail, GCP, Azure) without heavy CPU reliance.

Key Innovations:

1. High-Density FFN (HD-FFN): Unlike standard 1B models, JiRack-1.5B utilizes an 8192-dimension intermediate layer within ternary constraints, offering 3B-class semantic quality.
2. SWA Fusion (SwiGLU-Attention): A novel compute kernel that fused FFN and Attention passes, maintaining thermal stability < 80°C even under heavy load.
3. Buffered Routing Embedding (BRE): Minimizes data movement between HBM and compute units, specifically optimized for ROCm environments.

⚠️ Changes

Alpha v1.3 has RoPe fix. So high Quality response now Old model renamed to model_tag.safetensors . it is for RAG with many counter questions style . The model had trainings with Supervised fine tuning technics. It is sasy to use LoRa to fine tune as LLama 3.2 1b LoRa adapter .
Will do model triple compression for production soon

---

📊 Technical Specifications

Training & Fine-Tuning Efficiency

✅ Consumer GPU Ready: Fine-tuning via LoRA (r=8) requires only ~8 MB of additional weights. ✅ 70% VRAM Reduction: Fits on RTX 3060, GTX 1660 Ti, or high-end integrated GPUs. ✅ EASY TRAIN: So you do regular Fine-Tuning ( SFT or LoRa ) then use pack script to convert weights to ternary . So current weights in fp16 but during training model learn how live in ternary format

---

💬 Chat Demo ( Release + Google Turbo Quant)

Current Checkpoint: Feb 26, 2026 (Logic Verification Mode)

• (ai_env) root@jirack1:~/JiRackTernary1/new# python chat_jirack_ternary1b_pt_turbo_packed_v1.py

• --- 🖥️ Device: cpu (JiRack Turbo Engine) ---

• --- 🚀 Loading JiRack 1B (Auto-Quantizing Weights) ---

• ✅ Weights loaded. System will quantize them on the first pass.

• User: Hello

• JiRack: Hello! How are you?

• (Time: 0.97s on CPU)

• User: How are you dong ?

• JiRack: I’m doing fine, I’m not sure what you’re asking. What do you mean?

• (Time: 0.65s on CPU)

• User: What is capital of France ?

• JiRack: It’s a tiny city with a population of about 35,000 people. It’s in the center of France, and is a economic center for the French government. It’s also the capital of French-speaking France, and is home to the National Museum of Art and Natural Heritage.

• (Time: 1.88s on CPU)

• User: What is capital of France name ?

• JiRack: The capital of Paris is Paris.

• (Time: 0.34s on CPU)

• User: What is Sun ?

• JiRack: Sun is the largest planet in our solar system, which we call the Sun. It is 400 billion years old. It has 1.4 billion years in diameter, and 6.5 billion miles in diameter.

• (Time: 1.34s on CPU)

Note: The current .pt file is ~6 GB due to heavy embedded metadata for research tracking. The final production version will be cleaned and compressed to ~1.5 GB - 2.0 GB.

---

📂 Repository Contents

• JiRackTernaryPyTorch_1b.py: Core architecture file.
• invention_description.md: Detailed technical breakdown for patent examiners.
• performance_data.md: Benchmarks on ROCm/AMD hardware.
• NDA.md: Confidentiality agreement for commercial evaluators.

📧 Contact & Licensing

For joint venture opportunities, hardware integration, or licensing inquiries:

• Email: [email protected]
• Phone: +1 (516) 777-0945
• Location: New York, USA