We open-sourced Gepard 1.0 - a 555M streaming TTS that starts talking as text arrives.
~50ms to TTFA, ~20x RTF on 1 RTX 5090.
Voice cloning from seconds of audio.
vLLM native.
Apache 2.0.
Demo + weights:
https://t.co/H5MBzJxIRr
New blog post: How to Build a Diffusion Language Model.
Diffusion LLMs went from open problem to reality in 2 years (Mercury, Gemma Diffusion, Nemotron Diffusion). With my Cornell group, we wrote up the research advances that make them work.
We use today's OSS large diffusion models as examples and show how they're built from core techniques:
• masked diffusion (MDLM)
• iterative refinement (UDLM, ReMDM)
• variable-length generation (block diffusion, encoder-decoder architectures)
• controllable generation (D-CFG, D-CBG)
• fast samplers (Duo)
• RL post-training (d1, d2)
This post covers research led by a talented group of Cornell PhD students including @mariannearr@SchiffYair@Guanghan__Wang
The content is adapted from talks I gave this year on dLLMs. It covers most of the main ingredients behind open-source diffusion language models today.
Link: https://t.co/lpu3BCLwiE
A few days ago, Dr Hussam Abu Safiyeh (@dr.hussam73) said these words to his lawyer: “This is the last time you will see me...they brought me here to kill me. I don't see myself surviving. This is the end”.
https://t.co/IekOkgmTvg
i've had a pet technical problem for ten years. i know the solution space inside and out. a few months ago gpt-5.5 made a massive leap that broke new ground, but the solution was not elegant. last night, fable found the elegant solution 😳