README.md

---
title: AGILLM 4.3 GUI (CPU)
emoji: 🦊
colorFrom: blue
colorTo: green
sdk: gradio
app_file: app.py
pinned: false
license: mit
models:
  - OpenTransformer/AGILLM-4.3
---

# AGILLM 4.3 — Local Inference GUI (free CPU Space)

Web version of the AGILLM 4.3 desktop GUI. 1.2B-parameter research LLM trained
from scratch (DiffusionBlocks + MoE, AR/SAT/NAT decode heads).

- **NAT** (non-autoregressive mask-predict): fastest head; streaming shows the
  token canvas filling in confidence order — a live view of the parallel decode.
- **SAT var / SAT fixed**: semi-autoregressive block decode (gate-chosen vs
  fixed stride).
- **AR**: classic left-to-right decode, streams token by token.

The checkpoint is mid-pretraining — output is word-salad by design at this
stage. Free CPU is slow; see the ZeroGPU twin Space for fast generation.