Tyger: open-weights 0.6B model adds full-duplex turn control to voice agents
ylankgz · reddit · 2026-10-17
nineninesix released Tyger, an open-weights small model for full-duplex turn-taking in voice agents (Hugging Face: nineninesix/tyger):
- Base: NVIDIA Nemotron 3.5 ASR Streaming 0.6B + a 14M causal GRU; every 80ms with no lookahead it outputs one of three states: answer (user finished), stop (user interrupted, agent stops), or chime in (agent says "mm-hm" and keeps listening)
- Hears both the mic and the agent's own playback, so it won't mistake its echo for the user
- tyger-s2s (Apache-2.0) is a full voice agent stack: Tyger + Gemma 4 E4B FP8 + VoxCPM2, with tool calls, up to 4 concurrent conversations, and 0.7s reply latency on an RTX 5090; fully local by default with an optional OpenAI "expert" backend off by default
- Limitations: needs 28GB VRAM, tested only on 5090, FP32 only, offline benchmarks only
The team solicits feedback on turn-taking feel and 3090/4090 compatibility reports.
Related event: Open-source Tyger 0.6B model gives voice agents full-duplex turn control(2 posts)→
More from coding & agent
- Screenshot to spreadsheet in 11 seconds with one line of PowerShell — dfinke · 2026-10-17
- Deploy MiniCPM-V 4.7 locally by pointing Claude at its Hugging Face page — NielsRogge · 2026-10-17
- ThePrimeagen mocks AI coding: infinite tokens still produce an Electron app — seatedro · 2026-10-17
- Dev uses Unreal Engine MCP server with an agent to build a virtual workplace in an hour — Teknium · 2026-10-17
- Generative UI 101: Building Live Interfaces with Agents on Opus 5.5 and Rivals — EXM7777 · 2026-10-17
- MCP server review: gate tool writes on auth/traversal findings, not test score — SamuelFmedeiros · 2026-10-17