README.md

metadata

title: PROMETHEUS v1.0 — World Model Interactive Demo
emoji: 🔥
colorFrom: red
colorTo: blue
sdk: gradio
pinned: true
license: apache-2.0
tags:
  - world-model
  - embodied-ai
  - embodied-intelligence
  - 3d-simulation
  - cognitive-ai
  - wm-bench
  - final-bench
  - prometheus
  - vidraft
  - interactive-demo
  - real-time
  - game-ai
  - npc
  - skeleton-animation
  - glb
  - three-js
  - fastapi
  - benchmark
  - agi
  - proto-agi
  - aether
  - hephaestus
  - flooddiffusion
  - motion-generation
  - threat-response
  - emotion-escalation
  - korean-ai
  - k-ai
short_description: World-first embodied AI world model

🔥 PROMETHEUS v1.0 — World Model Interactive Demo

"The world's first World Model to pass WM Bench Track C verification"
WM Score 726 / 1000 · Grade B · 47 FPS · RTX5070 local

▶ Launch Demo  |  🏆 Leaderboard  |  📦 Dataset

---

What is PROMETHEUS?

PROMETHEUS v1.0 is a real-time, interactive 3D World Model that you can experience directly in your browser — no installation required. It is the official baseline model of WM Bench, the world's first benchmark measuring the cognitive intelligence of world models.

Unlike conventional world models that only generate realistic-looking video, PROMETHEUS actively perceives, predicts, judges, and acts in a dynamic 3D environment. It currently ranks #1 on WM Bench among 26 registered world models including models from Meta, NVIDIA, Google DeepMind, and OpenAI.

---

Quick Facts

WM Score	726 / 1000
Grade	B (threshold: ≥600)
Track	C — Live Demo + Official Verification ✓
FPS	47 FPS (RTX5070, local/16GB)
GPU (Space)	T4 (HuggingFace)
Architecture	AETHER + PROMETHEUS + HEPHAESTUS
Motion Engine	FloodDiffusion-VIDRAFT
Joints	263-joint skeleton with GLB retargeting

---

Architecture: The Three Sacred Fires (三神器)

AETHER     — The Fire of Wisdom (Proto-AGI Brain)
  Open-architecture: plug any LLM as the cognitive engine
  Emergence · Meta-cognition · SLAI self-evolution · Multi-agent

PROMETHEUS — The Fire of Perception (World Model Engine)
  Perception → Prediction → Judgment → Action loop
  FloodDiffusion-VIDRAFT motion generation

HEPHAESTUS — The Fire of Form (Body Engine)
  263-joint skeleton · GLB retargeting
  Humanoid · Tank · Drone body support

---

WM Bench Results (Track C — Officially Verified)

Pillar	Score	Max	Key Categories
👁 P1 Perception	140	250	C01 Env.Awareness: 65 · C02 Entity Recog: 75
🧠 P2 Cognition	390	450	C04 Threat Diff: 90 · C03 Prediction: 85 · C05 Emotion: 85
🔥 P3 Embodiment	196	300	C09 Realtime: 85 · C08 Motion-Emotion: 80 · C10 Body-Swap: 35
Total	726	1000	Grade B

Strengths

• C04 Threat-Type Differentiation (90/100) — Clearly distinguishes beast vs. human threats and responds proportionally
• C05 Autonomous Emotion Escalation (85/100) — No prior research exists for this category; PROMETHEUS achieves 85
• C03 Prediction-Based Reasoning (85/100) — 4-directional future danger state prediction

In Development

• C10 Body-Swap Extensibility (35/100) — Cross-embodiment transfer (Phase 3 roadmap)

---

How to Experience

1. Open the Space above
2. Select a world: Castle / Inferno / Horde / Countdown / Dilemma
3. Control the character with keyboard/mouse
4. Observe how the AI perceives threats, predicts futures, and responds with emotionally grounded motion
5. Try Tank Mode — the same cognitive brain drives a completely different body form

No GPU required on your end. Runs on T4 in the cloud.

---

Frequently Asked Questions

Q: What makes PROMETHEUS different from other world models like Genie 2 or Cosmos?
A: Most world models (Genie 2 ~449, Cosmos ~498 on WM Bench) are optimized for visual realism (FID/FVD). PROMETHEUS is designed to score on cognitive intelligence — how it judges, remembers, and emotionally responds. On WM Bench Cognition (P2), PROMETHEUS scores 390/450, significantly ahead of current estimates for other models.

Q: What is WM Bench?
A: WM Bench is the world's first benchmark measuring the cognitive abilities of world models. It evaluates across 3 pillars (Perception · Cognition · Embodiment), 10 categories, and 100 scenarios. 6 of the 10 categories are defined here for the first time in any benchmark. See WM Bench Leaderboard.

Q: Can I submit my own world model to WM Bench?
A: Yes. Download the dataset, run your model on 100 scenarios, and submit results to the Discussion board. Track A (text-only API) requires no special hardware.

Q: What is the FINAL Bench Family?
A: FINAL Bench measures text AGI (metacognitive reasoning), WM Bench measures embodied AGI (world model cognition). Together they form the FINAL Bench Family by VIDRAFT. FINAL Bench reached HuggingFace global dataset Top 5 and was covered by 4 press outlets in February 2026.

Q: What is AETHER?
A: AETHER is the cognitive brain layer of PROMETHEUS — an open-architecture Proto-AGI system that accepts any LLM as its reasoning engine. It features emergence, meta-cognition, SLAI self-evolution, and multi-agent coordination (synergy/critique dynamics).

---

Related Resources

Resource	Link
🏆 WM Bench Leaderboard	https://huggingface.co/spaces/FINAL-Bench/worldmodel-bench
📦 WM Bench Dataset	https://huggingface.co/datasets/FINAL-Bench/World-Model
🧬 FINAL Bench (Text AGI)	https://huggingface.co/datasets/FINAL-Bench/Metacognitive
🏆 FINAL Bench Leaderboard	https://huggingface.co/spaces/FINAL-Bench/Leaderboard
📊 ALL Bench Leaderboard	https://huggingface.co/spaces/FINAL-Bench/all-bench-leaderboard

---

Citation

@misc{prometheus2026,
  title   = {PROMETHEUS v1.0: An Embodied World Model for Cognitive Intelligence Evaluation},
  author  = {Kim, Taebong},
  year    = {2026},
  url     = {https://huggingface.co/spaces/FINAL-Bench/world-model},
  note    = {WM Bench Track C Verified · Score 726/1000 · Grade B}
}

License: Apache 2.0

---

Part of the FINAL Bench Family by VIDRAFT · "Beyond FID — Measuring Intelligence, Not Just Motion."

#WorldModel #EmbodiedAI #WMBench #FINALBench #PROMETHEUS #VIDRAFT #CognitiveAI #AGI #ProtoAGI #AETHER #3DSimulation #RealtimeAI #KoreanAI #HuggingFace #BenchmarkAI #MotionGeneration #NPC #GameAI #EmbodiedIntelligence #FloodDiffusion