# Kyutai — Diffusable Latents from Structure-Agnostic Distillation

- Company: Kyutai (kyutai.org)
- Announced: 2026-09-28
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://kyutai.org/blog/2026-09-28-structure-agnostic-distillation
- Record: https://forck.live/items/18334-diffusable-latents-from-structure-agnostic-distillation
- Subject: Moshi / Pocket TTS

Kyutai's research shows that a single vector per image, obtained by pooling, is sufficient to make a latent representation diffusable regardless of its original structure.

## Evidence

Verbatim from https://kyutai.org/blog/2026-09-28-structure-agnostic-distillation:

> Pool, then distill: one vector per image is enough to make a latent diffusable, whatever its shape.

---

Record: https://forck.live/items/18334-diffusable-latents-from-structure-agnostic-distillation
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md