Research weights arrive with different controls, inputs, and runtime assumptions. Windflow preserves those differences while giving every model the same typed session boundary: commands in, tracks out.
10
Public catalog
Model runtimes
09
Production
Published runtime status
01
Session API
One realtime protocol
02
Developer SDKs
JavaScript and Python
Credits / second
17
14B Diffusion Transformer, autoregressive chunked diffusion
One continuous, steerable clip on the main_video track
Prompt · Image
Resolution
Negotiated per session
Playout
Negotiated per session
Chunk
33 native frames in the production profile
Credits / second
21
Autoregressive multi-shot DiT (Wan2.2-TI2V backbone)
A multi-shot clip on main_video; chunk_index resets per scene while session_chunk keeps counting
Prompt
Resolution
1280 × 704
Chunk
29 frames
Credits / second
19
2B streaming editing Diffusion Transformer
The edited stream on main_video, with original motion preserved
Prompt · Video · Webcam
Resolution
1280 × 704 native edit and delivery
Chunk
24 frames
Credits / second
21
16B multimodal DiT, chunk-autoregressive, two-step distilled
The edited stream on main_video, with source motion and identity preserved
Prompt · Image · Video · Webcam
Resolution
840 × 480 interactive tier; 1248 × 720 detail tier
Chunk
8 frames
Credits / second
24
Seed-image conditioned navigable world model
A navigable stream on main_video you steer in real time
Prompt · Image · Movement
Resolution
1408 × 768 delivery (704 × 384 generation)
Chunk
Movement and look apply at the next generation-block boundary
Credits / second
24
Real-time AR world model (SkyReels-V2 base, Self-Forcing distilled, 3-step)
A navigable generated world stream on main_video
Image · Movement
Resolution
640 × 352
Chunk
12-frame AR blocks; actions land at the next block
Credits / second
23
5B controllable world generation model (Wan2.2-TI2V backbone)
A controllable world rollout on main_video
Prompt · Image · Movement
Resolution
704 × 1280 portrait in the production runtime; negotiated per session
Chunk
Action-sequence conditioned rollout
Credits / second
23
5B keyboard-conditioned world model (Wan2.2-TI2V backbone, LongForcing causal student)
A third-person navigable world stream on main_video
Prompt · Image · Movement
Resolution
1344 × 768 delivered (896 × 512 generation + learned super-resolution)
Chunk
12-frame AR blocks; held keys are sampled at each block boundary
Credits / second
150
35B MiniMax H3 text-to-video-and-audio model, DMD2-distilled to four DiT calls with 90% sparse VSA-H3 attention
Video on main_video and generated audio downmixed to mono on main_audio, aligned per window
Prompt
Resolution
960 × 544 on the current 4 × H100 realtime profile
Playout
24
Chunk
345 frames per window (14.38 s at 24 fps; published as the 15 s profile)
Credits / second
40
22B joint audio+video DiT, two-stage, generated one window at a time
Video on main_video and the audio generated with it on main_audio, aligned per window
Prompt · Image
Resolution
832 × 448 measured; the canvas is configurable in multiples of 64
Playout
24
Chunk
241 frames per window (10.04 s)
Model contract
