Block unveils Dessa: hyper-realistic speech-to-speech model trained on a single GB200 NVL72
altryne · x · 2026-10-24
- A small research team inside Block has unveiled Dessa, a speech-to-speech mode so realistic it's said to break through another uncanny valley wall.
- Jack Dorsey says it runs on his MacBook (small model) and was trained on a single GB200 NVL72 GPU.
- Not yet publicly available — only a waitlist and demo reels from team member Ragavan Thurair.
Related event: Block's Dessa: Voice-Trained Full-Duplex Speech Model Runs on a MacBook(8 posts)→
More from Multimodal
- Community fine-tune fixes Qwen-Image 2.1 VAE 8-pixel grid artifact — Crazy-Repeat-2006 · 2026-10-24
- Character LoRA trained on 21 images fails across models—training setup questioned — IntrovertedOtaku03 · 2026-10-24
- Tencent's Hy-Nomas video world model ported to native ComfyUI, runs on 12GB VRAM — LatentSpacer · 2026-10-24
- Video models replace NeRF/3DGS pipelines: single-frame novel view synthesis, open-source soon — jonstephens85 · 2026-10-24
- Recreating PS2-Era Harry Potter With FastH3-V2 in 4 Steps — aziib · 2026-10-24
- A Full AI Music Video Made Locally on DGX Spark With MiniMax H3 — Ornery-Block-3522 · 2026-10-24