ternary-weights

From-scratch 0.15B BitNet 1.58-bit (ternary) RWKV-7 LLM trained on an 8GB GPU — QAT, 2-bit packing, RFT, local chat app.

Ternary Weights Network for Darknet

TQ1.0 ternary inference engine for BitNet b1.58 on CPU. Pack + run Falcon3-1B/3B/7B/10B, no GPU needed.

Pre-training research stack for a 3.67B BitNet b1.58 + 8-expert MoE + Liquid/CfC LLM — architecture, data pipeline, knowledge distillation, eval harness and closure gates. The canonical model is NOT yet trained; no capability or benchmark claims are made.

Pure-Julia CPU inference engine for BitNet b1.58 ternary LLMs

PERSPECTIVE v2 — A 1.05 trillion parameter sparse Mixture-of-Experts language model that runs on consumer hardware (4 GB VRAM + 32 GB RAM). Features O(1) perspective decay recurrence, 3D torus manifold routing, native ternary {-1,0,+1} weights, holographic distributed memory, and hard geometric safety constraints. Built in Rust.

A transpiler written in Rust aiming to target a specialised language for ternary model training and optimised hardware constraints such as IoT/embedded devices

Experimental ternary (1.58-bit) backend for PyTorch.

Interactive Training Dashboard & CAGS-Operator Verification for JamOne Nano.

Mycelium: a verification-first spiking network derived from the CeliumNeUR chip constraint set — bit-exact to silicon-verified RTL, ternary block-sparse, learned-then-frozen topology

Custom CUDA kernels for accelerating 1.58-bit ternary LLM inference with 2:4 structured sparsity on consumer Ampere GPUs. Exploits both ternary arithmetic (no multiplies) and hardware sparse tensor cores to maximize throughput on RTX 3060. Based on the Sparse-BitNet paper (Zhang et al., 2026).

First 1-bit (BitNet b1.58) recursive reasoner for Sudoku-Extreme - distilled from a 7M-param FP TRM teacher into a 1.4 MB ternary student