DeepSeek V4.1 Flash
Native generation across code, problem-solving, and storytelling.
INT21 / Inference Engine Factory Built by Swarm
Inference Engine
Factory.
Built by Swarm. Produced at scale.
From model to native inference engine.
Watch the factory's output across language, images, video, and sound.
20Engines
7Model types
3GPU families
The collection
See the engines run.
Hover to preview. Click for sound.Tap a film to watch with sound.
20 engines · 7 model types
Recorded on real hardwareNo engines found.
Try another model, capability, or GPU.
01
Language
8Code, reason, write. Watch the tokens arrive.
DiffusionGemma
An API, SQL analytics, and a product launch plan.
GLM 5.3
Python, GPS relativity, and science fiction at inference speed.
GLM 5.3 Flash
A Flash runtime, three tasks, and live tokens per second.
HY4 Preview
Async code, a factory schedule, and a weather-bending story.
Kimi K3
Three live requests. Code, reasoning, and creative writing.
MiMo V2.6 Pro
Fresh prompts, streamed responses, and measured decode speed.
Qwen3.8 Flash Next
Build a cache, plan a launch, and discover a lost spacecraft.
02
Image
2From a prompt to the finest detail. Generate and edit in 1K and 2K.
03
Video
2Build moving worlds. Generate scenes with sound.
04
Speech
3A whisper, a new voice, a different delivery. Hear the result.
05
Music
1Turn the same words into a completely different sound.
06
Transcription
2Speech becomes text, across languages and styles.
07
Vision
2Find objects. Read documents. Make the visual world usable.
Films show selected inference runs. Performance figures apply to the hardware and settings shown in each demo.
The factory behind the films
Use compute
to improve compute.
Swarm builds the inference stack, tests it on target GPUs, and carries the results into the next iteration. Every model is another engine from the factory.
Inside the factory