The AI stack was built for language models. We’re rebuilding it for world models.
AI is entering a new era. Models are evolving from understanding and generating text to understanding and simulating the world itself—its geometry, physics, causality, dynamics, and interactions.
While the models have changed, the infrastructure has not. Today’s training and inference stack was designed for autoregressive language models. World models process and generate images, video, audio, actions, and 3D environments, demanding orders of magnitude more compute and fundamentally different execution patterns.
We are a team of experienced engineers and researchers with a track record of building world-class AI infrastructure at Fireworks AI, Verkada, Snowflake, Meta, Google, Alibaba, Apple, and Microsoft.
MiniMax H3 / 720p / 5s Video:
142 sec baseline → 13 sec on TensorScale)The TensorScale API (https://api.tensorscale.io) serves hosted video and image generation models over HTTPS:
| Model | Model ID | Tasks | Synchronous | Streaming | Asynchronous |
|---|---|---|---|---|---|
| MiniMax H3 | MiniMax-H3 |
Text, frames and references | Yes | Yes | Yes |
| LTX-2.5 Fast | ltx2-5-fast |
Text and image to video | Yes | Yes | Yes |
| Wan 3.0 | wan-3.0 |
Generation and editing | — | — | Yes |
| MiniMax H3 Image | MiniMax-H3-Image |
Generation and editing | Yes | — | Yes |
| Qwen Image 2.1 | Qwen-Image-2.1 |
Generation and editing | Yes | — | Yes |
| MOSS Image | MOSS-Image-Base, MOSS-Image-SFT |
Text to image | Yes | — | Yes |
| SenseNova U1.5 | SenseNova-U1.5 |
Generation and editing | Yes | — | Yes |
| FLUX3 Image Klein | FLUX3-Image-Klein |
Generation and editing | Yes | — | Yes |
curl
curl "https://api.tensorscale.io/v2/ltx-2.5/fast" \
--max-time 7200 \
-H "Authorization: Bearer ${TENSORSCALE_API_KEY}" \
-H "Content-Type: application/json" \
-H "Accept: video/mp4" \
-d '{
"prompt": "A cinematic aerial push toward a lighthouse as waves roll against the cliffs.",
"width": 1280,
"height": 768,
"num_frames": 121,
"frame_rate": 24,
"seed": 10
}' \
--output "ltx2-5-fast.mp4"
Python
import os
from pathlib import Path
import requests
resp = requests.post(
"https://api.tensorscale.io/v2/ltx-2.5/fast",
json={
"prompt": "A cinematic aerial push toward a lighthouse as waves roll against the cliffs.",
"width": 1280,
"height": 768,
"num_frames": 121,
"frame_rate": 24,
"seed": 10,
},
headers={
"Authorization": f"Bearer {os.environ['TENSORSCALE_API_KEY']}",
"Accept": "video/mp4",
},
timeout=7200,
)
resp.raise_for_status()
Path("ltx2-5-fast.mp4").write_bytes(resp.content)
Codex, ChatGPT, Claude, Cursor, Gemini CLI, and Antigravity can follow one trusted setup skill to configure Streamable HTTP and OAuth—without putting API keys in chat:
Read https://tensorscale.io/setup-mcp-skills.md and configure TensorScale MCP for me.