metadata
license: mit
language:
- en
library_name: generic
tags:
- llm
- local
- numpy
- uncensored
- multi-agent
- skill-creation
- conversation-mesh
- self-improving
- voice-assistant
- jarvis
- splitbit
- 100-local
- zero-limitation
- always-on
- autonomous
pipeline_tag: text-generation
SplitBit LLM v1.0.0
A fast, uncensored, 100% local LLM built from scratch with SplitBit sub-byte token encoding, recursive link memory, multi-agent skill creation, conversation mesh, and zero-limitation tool calling.
Key Features
- Pure NumPy — no PyTorch dependency, runs on any machine with Python 3.10+
- SplitBit tokens — sub-byte token encoding with 4-20x memory compression
- 100% local — no API calls, no cloud, no external dependencies
- 5 AI agents — planner, coder, researcher, reviewer, executor
- Conversation mesh — all agents converse with each other to build skills
- Skill cascade — building a skill auto-generates related skills
- Always-on daemon — 24/7 autonomous operation with 100-project mode
- Fast reply cache — near-instant responses with 3-layer caching
- Zero-limitation tools — full terminal control, code editing, self-modification
- Mass storage vault — auto-resizing storage with compression and cleanup
- First-run naming — "Hi, I am an Incentives Inc. LLM. What would you like to name me?"
- Built-in Jarvis — voice assistant with wake word, STT, TTS
- Image generation — 100% local NumPy-based procedural images
Installation
pip install splitbit-llm
Or from source:
git clone https://github.com/incentivesinc/splitbit-llm.git
cd splitbit-llm
pip install -r requirements.txt
Quick Start
# First run — it asks you to name it
python -m splitbit_llm chat
# Jarvis voice assistant
python -m splitbit_llm jarvis
# Start web server with REST API
python -m splitbit_llm serve
# 24/7 autonomous daemon mode
python -m splitbit_llm daemon
Hardware Tiers
Auto-detects your hardware and adjusts model size:
| Tier | RAM | Model | Quantization |
|---|---|---|---|
| Mobile | <2GB | 2L/4H/d128 | Ternary (1.6bpw) |
| Minimal | <4GB | 3L/4H/d256 | Q2_K (2bpw) |
| Light | <8GB | 4L/8H/d384 | Q3_K (3bpw) |
| Standard | <16GB | 6L/8H/d512 | Q4_K (4bpw) |
| Full | <32GB | 8L/16H/d768 | Q5_K (5bpw) |
| Maximum | <64GB | 12L/16H/d1024 | Q8_0 (8bpw) |
Architecture
- Model: Custom transformer with SplitBit token encoding (pure NumPy)
- Memory: SQLite-backed episodic + semantic memory with recursive links
- Agents: 5 specialized AI agents with autonomous conversation mesh
- Skills: Auto-extracted from conversations, pooled by category, cascaded
- Tools: 9 zero-limitation tools including terminal control and self-modification
- Storage: Auto-resizing mass storage vault with compression tiers
- Server: FastAPI REST API with 20+ endpoints
License
MIT License — Copyright (c) 2026 Incentives Inc.
Author
Built by Incentives Inc.