---
title: SAM-MM Multimodal Demo
emoji: 🧠
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.18.0
app_file: app.py
pinned: false
license: other
short_description: 58M offline multimodal model for vision and physics
---

# SAM-MM — multimodal reasoning demo

Interactive demo of **SAM-MM** (AMFORGE), a ~58M-parameter offline multimodal model that
perceives synthetic frames (and a log-mel spectrogram for audio scenes) and answers with
either a `[CHAT]` reasoning trace or a `[ACTION]` JSON record.

Pick a scene; each one is **freshly generated** and the model decodes token-by-token —
nothing is hard-coded. Physics and motion are the model's strength (its latent world-model
carries real dynamics); OCR generalizes to unseen numbers; the cross-modal action family
is weaker; **audio is the weak modality** (trained on synthetic pseudo-mel).

## Setup

The model weights live in **private** HuggingFace repos. Add a Space **secret** named
`HF_TOKEN` (a token with read access to the `AMFORGE/sam-mm-*` repos). The app pulls the
checkpoint and tokenizer from the Hub on first run.

Architecture internals, tokenizer construction, and data generators are proprietary and
not exposed by this demo.