Model Hub

Janus-Pro-7B unified multimodal (image generation). Validated engine surface: PyTorch branch (compiled + sequential, byte-identical); image-autoregressive on the Triton engine is pending and refused with a clear message until then.

Image-to-video generation at native 720x1280, 88-frame clips conditioned on a first frame.

Swin2SR Classical ×4 — Swin Transformer V2 image super-resolution model that upscales images 4× while restoring fine detail; the classical-SR variant for clean low-resolution inputs. Apache 2.0.

Flex.1 Alpha — Ostris 8-billion-parameter rectified-flow text-to-image model derived from FLUX.1-schnell, featuring a trainable guidance embedder that makes it highly flexible to fine-tune. Apache 2.0.

High-motion text-to-video generation, asymmetric diffusion transformer, 10B parameters.

Qwen3-Omni 30B-A3B Instruct — quad-modal understanding (text, image, audio, video) + generative speech output (talker + MTP mono-step predictor + codec vocoder) on one build; thinker branch, MoE 128-expert

OpenAudio S1-Mini — Fish Audio compact multilingual (13-language) text-to-speech model offering expressive, emotion-controllable voice synthesis and zero-shot voice cloning. Non-commercial license (CC-BY-NC-SA-4.0).

Parakeet TDT 1.1B — NVIDIA high-throughput speech-to-text model (FastConformer encoder with a Token-and-Duration Transducer) for fast, accurate English transcription with punctuation, capitalization and timestamps. CC-BY-4.0.

MiniCPM-o 4.5 — omni build: tri-modal understanding (text, image, audio) plus generative voice output (speech codec AR + flow-matching token2wav vocoder, 24 kHz); SigLIP NaViT + perceiver resampler, whisper-medium audio, Qwen3-8B dense LM

All-in-one video creation and editing (VACE), 1.3B parameters. Text-to-video generation is validated and reproducible: at 480x832, 9 frames, guidance 5.0 it renders bit-identically to its 2026-06-28 validation on both engines (re-checked 2026-09-03). Declared limitation at this release: generation conditioned on a REFERENCE IMAGE is not supported - that path stops on a channel-size mismatch and is tracked as a named fix. Use the prompt-only mode; the sibling Wan2.1-T2V-1.3B covers plain text-to-video as well.

SwinIR Classical x2 - Swin Transformer image-restoration model for 2x super-resolution, reconstructing sharp high-resolution detail from low-resolution images. Works at any input size. Apache 2.0.

Swin2SR RealWorld x4 - Swin Transformer V2 super-resolution trained on BSRGAN-style degradations, for real-world photographs carrying noise, blur and compression artifacts rather than clean sources. PSNR-oriented variant, so it favours fidelity over invented detail. Upscales 4x. Apache 2.0.