Model Hub

Canary-Qwen 2.5B — NVIDIA speech-augmented language model pairing the Canary (FastConformer) audio encoder with a Qwen LLM for English speech recognition plus LLM-style summarization and question-answering over transcribed audio. CC-BY-4.0.

DeepSeek-MoE 16B chat — fused-MoE runtime, greedy decode contract

Voxtral Mini 3B — Mistral AI audio-language model built on Ministral 3B for speech transcription and spoken-audio understanding, handling long-form, multilingual audio with built-in question-answering over what it hears. Apache 2.0.

Orpheus 3B — Canopy Labs LLM-based text-to-speech model built on Llama 3 architecture, expressive natural speech generation. Apache 2.0.

VibeVoice 1.5B — Microsoft text-to-speech model for long-form, multi-speaker audio such as conversational podcasts, producing expressive voices with natural turn-taking between speakers. MIT license.

IBM Granite 3.1 1B-A400M Instruct — a 32-expert mixture-of-experts language model (8 experts per token, 24 layers), the first MoE container through the fused expert dispatch; traced and validated on NVIDIA V100 in sequential, compiled and Triton modes

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

HAT-S x4 - XPixel Group Hybrid Attention Transformer (Small) super-resolution model that activates more input pixels through combined channel and window attention for 4x image upscaling. Works at any input size. Apache 2.0.

Efficient 720p text-to-video generation, 2B parameters, linear-attention diffusion transformer.