Gen 2 Deterministic AI - Tribot-9.98m-micro Public Release

Models
1

Tribot-9.98m-micro: Public Model Release Overview

Tribot-9.98m-micro is a foundational language model developed by Triskel Data Pty Ltd. It is the result of a fully independent pipeline, built from the ground up with no external model dependencies. Every stage, from data cleaning to architecture and training, was designed for determinism, auditability, and ultra-portability.

What It Is

Tribot-9.98m-micro is a compact transformer model trained on a cleaned and structured 2025 Reddit comment dataset. It contains 9.98 million parameters and a 153152-token vocabulary, all structured to support offline, deterministic question answering and conversational tasks. It is designed to run efficiently on CPUs, edge devices, and embedded systems.

This is not a toy. It is a real, losslessly trained foundational AI model that operates entirely offline with no cloud requirements, no telemetry, and no reliance on external APIs.

What It Contains

The public package includes:

  • model.pt: The trained PyTorch model file
  • vocab.json: The full vocabulary (case sensitive)
  • inference settings.py: A simple Python script for local testing
  • README.md: Installation and usage overview
  • LICENSE.md: Detailed license structure and pricing
  • model_card.md: Architecture, dataset, and limitations summary
  • inference_instructions.md: Setup and usage guide
  • screenshots/: Two images demonstrating training and inference working locally

Training Details

  • Trained on: Structured Reddit dataset (2025)

  • Tokens used: 55GB+ of cleaned text

  • Vocabulary: 153152 words (full upper and lower case)

  • Architecture: Custom MiniTransformer

    • Layers: 2
    • Heads: 8
    • d_model: 32

  • Training strategy: One-pass, lossless deterministic training

  • Final model size: 127MB uncompressed

  • Compressed package: 62MB (7z format)

Performance and Deployment

Tribot-9.98m-micro runs entirely offline. It is suitable for devices like:

  • Laptops
  • Desktops
  • Intel N100 mini PCs
  • Raspberry Pi 4 and 5
  • ARM Cortex-A76 class smart TVs
  • Other consumer-grade CPUs with at least 2GB of RAM

With quantization and minor optimization, it can run on systems with minimal resources. Real-time interaction is achievable on mid-range and better devices. The model can support local voice assistant behavior, chatbots, or command processors in offline smart systems.

It is not suited for microcontrollers or devices with less than 1GB of RAM. Extremely low-end TV sticks or legacy devices will not perform well unless hardware is upgraded.

Inference Pipeline

Tribot’s inference setup is minimal. Install Python 3.8 or newer, install PyTorch 2.0+, place the files in one folder, and run the included inference settings.py. The model responds to prompts locally and deterministically.

No internet is required. No background processes. No data leaves the machine.

Licensing Overview

Tribot-9.98m-micro is licensed under a four-tier commercial structure:

Tier 1: Developer (Free)
Use: Non-commercial, personal, and research only
Restrictions: No commercial use, resale, or redistribution

Tier 2: Startup / Nonprofit
Use: Internal use only, up to five seats
Price: 500 USD one-time license

Tier 3: Commercial
Use: Unlimited internal use, commercial and SaaS integration with attribution
Price: 2500 USD one-time license

Tier 4: Enterprise / OEM
Use: Full redistribution, embedding, OEM, SaaS resale
Price: 100000 USD one-time license (custom contract)

All paid licenses are perpetual and royalty-free. No support or updates are included unless negotiated.

Key Differentiators

  • Fully foundational: No base model, no fine-tune, no third-party dependencies
  • Deterministic: Same inputs yield same outputs
  • Portable: Only 127MB total, compressible to 62MB with 450 to 1 ratio
  • No reliance on cloud, APIs, or online lookups
  • Private: Nothing leaves the device
  • Real-time: Can run usable local inference on mid-range consumer CPUs

Limitations

  • Not intended for legal, medical, or critical advisory applications
  • Knowledge is frozen at 2025 Reddit data
  • Limited reasoning capacity due to compact parameter count
  • No plugins, no real-time web knowledge

Suggested Use Cases

  • Voice assistants for smart TVs, remotes, and offline appliances
  • Embedded bots for diagnostics or device control
  • Academic tools or portable teaching assistants
  • Developer prototypes and experimental LLM stacks
  • Offline command systems in secured environments

Final Notes

This release shows that foundational AI does not have to be expensive, large, or reliant on cloud infrastructure. Tribot-9.98m-micro is a real, working, local AI engine that fits in a 62MB file and runs without any dependencies beyond Python and PyTorch.

You can copy it, share it, email it, or embed it in your hardware product. There are no runtime permissions, and no network choke points. Just drop it in and go.

*** this model is now obsolete, i removed the download link ***

Licensing: licensing@triskeldata.au
Support: support@triskeldata.au
Download: https://triskeldata.au

Triskel Data Pty Ltd
ABN: 83 660 229 961
ACN: 686 914 376

6

This is honestly fascinating—didn’t expect to come across a public release like Tribot-9.98m-micro so soon. Curious though, has anyone here tried comparing it with locally run models like llama cpp? I’ve seen a few folks mention it for lightweight inference, and wondering if something like that could be integrated or even benchmarked side-by-side.

7

Sorry, that one is very out dated. I’ve released far superior models on my website https://triskeldata.au The dev licenses are free. I’ve got 2 uploaded already. I’ll have another 3 more models ready by tomorrow.