Skip to content
AIBRAINWORX / OpenSourceCoder.comLOCAL-FIRST. BUILT IN THE OPEN.

SOURCE-REVIEWED MODEL PROFILE

AURA-5B

Audio-driven avatar research checkpoint; optimized H100 throughput is below its25fps playback setting.

Kyutai

Metadata reviewed 2026-11-20 · Original creator: Kyutai

PARAMETERS5.7B merged DiT per card
CONTEXTSee task-specific source settings
LICENSESnap Inc. Non-Commercial License
MODALITYPortrait image and speech to video

What it does

  • Four-step causal streaming student
  • Rolling KV cache
  • Non-commercial research terms

Local fit

Publisher:9.7fps end-to-end720p/H100 optimized pipeline;no consumer-GPU test.. Do not transfer the1.3B real-time claim to this variant. Merged weights do not include every encoder.

Video streaming uses a rolling cache;agent adapter context depends on the base and runtime. No independent long-context test.

Evidence and limits

Source reviewed;no local inference,hardware fit or benchmark test.

Publisher reference workflow;not independently tested

License and source

Artifact LICENSE reproduces Snap non-commercial research terms; separate components retain their licenses.

Read the original model card ↗

Reviewed source revision: 4442f4e4d59eb912ac09b8d1cc9d52522ee9b0d4

Pass the signal.