Model Hub

Qwen3-VL 30B-A3B Thinking — text, image and video understanding; 128-expert MoE, DeepStack vision, interleaved M-RoPE; multi-GPU sharded

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.

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

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

Ming-Lite-Omni-1.5 omni understanding (text + image + audio) AND generative image output (--mode image): windowed-ViT vision tower, varlen whisper audio encoder, multi-gate MoE decoder, in-package flow-matching diffusion stack (condition bridge + DiT denoiser + VAE) — 4-mode validated on both paths, seeded byte-reproducible generation