Kyutai
Moshi
Current
1.5
shipped 1mo · 8 Oct 2026
First real-time native speech-to-speech multimodal dialogue model with 160ms latency.
Specification
- Latency
- 160ms TTFB
- Full Duplex
- Native overlap & interruptions
- Weights
- Apache 2.0
- Origin
- France (EU)
Modalities: voice, audio, text
Moshi — 6-Axis Capability Radar
HEURISTIC ESTIMATEThese scores are rough estimates from category presets, catalog specs, and keywords in release notes. They are not measured benchmarks or test results. Catalog prices and context windows may refer to a different model variant.
Hover or tap an axis to inspect its estimate
Shipping Cadence & Next Drop Forecast
This is an uncalibrated estimate from limited catalog history. A ratio of 100% means one estimated interval has elapsed; it is not a probability. Next-release dates and versions remain unconfirmed.
827 days
±7d variance34 days
Shipped on 2026-10-080.4 drops/yr
1 observed intervals · limited history2029-01-12
Next version: UnannouncedOptional AI cadence commentary
Summarize the recorded history and its limitations. Future releases remain unconfirmed.
Model Version Capability Delta
AI-synthesized architectural jumps, parameter scaling, and capability shifts across releases.
Ready to analyze Moshi v1.0 vs v1.5
Click Synthesize Diff to analyze latent space adjustments, resolution leaps, parameter shifts, and migration caveats via Gemini 3.8 Flash.
Version history
2 releases tracked
1.5
minor8 Oct 2026 · 1moMoshi 1.5 drops latency to 110ms with multilingual conversational turn-taking and emotional mirroring.
- Native speech-to-speech dialogue with instantaneous human interruption handling.
- Open weights and Mimi audio codec updated with enhanced voice timbre clarity.
Recorded announcementSpecific source link · unverified ↗1.0
major3 Jul 2024 · 2yKyutai open-sources Moshi, achieving conversational speech-to-speech without STT/TTS cascade.
- First native audio-to-audio foundation model with realistic emotional prosody and backchanneling.
- Full duplex: can listen and speak simultaneously without audio cutoffs.
- Weights and Mimi neural audio codec released under Apache 2.0.
Recorded announcementSpecific source link · unverified ↗