personaplex-native-tool-use

PersonaPlex Native Tool Use — v169 Timer Proof

This repository contains a compact research adaptation checkpoint for nvidia/personaplex-7b-v1. It adds native action and environment lanes to a PersonaPlex/Moshi-style streaming state so the speech model can emit a typed tool request, observe the executor result in the same continuing state, and generate its own spoken confirmation.

The checkpoint is not a standalone model. You must separately obtain the PersonaPlex base weights from NVIDIA and accept their model license.

Code and reproducibility material: lazybutai/personaplex-native-tool-use

Verified narrow result

The promoted frozen test is a causal WAV replay of the request:

Count down two minutes for me.

With the pinned v169 checkpoint and seed, the persistent model session produced:

• exactly one native timer.create action;
• arguments {"minutes":"MIN_2"};
• action issued at frame 20;
• an allow-listed executor scheduling a real 120-second timer;
• one terminal OK result visible to the continuing model at frame 25;
• PersonaPlex-generated text: Sure, set a 2 minutes.;
• finite native assistant audio (RMS 0.016331, peak 0.442787).

The Windows eager run processed 46 frames at a measured real-time factor of 12.991. This establishes a reproducible tool-mechanics and same-state result-awareness proof. It is not a real-time performance claim.

What is native about it

The verified trajectory does not use an ASR transcript, external intent classifier, planner-generated tool JSON, host-authored confirmation, or external TTS. The action packet is emitted from learned model lanes. The typed executor observation is injected into the continuing causal model state, after which the model generates assistant text and audio.

Checkpoint

The checkpoint contains the compact agent-lane adaptation, not the PersonaPlex 7B base checkpoint.

Intended use

This release is intended for:

• research into agentic full-duplex speech models;
• inspection of native action/environment lane mechanics;
• reproducing the frozen two-minute timer trajectory;
• extending the action vocabulary and building broader training corpora.

Limitations

• This is a narrow, seed-pinned research checkpoint, not generalized tool use.
• The promoted result uses controlled causal WAV replay, not a live microphone.
• Tool behavior outside the trained timer trajectory is not established.
• Sampling stability remains limited.
• The Windows path is approximately 13 times slower than real time.
• A continuous low-latency conversation requires further Linux runtime work, negative/no-tool evaluation, microphone testing, and broader training.
• The generated sentence has rough grammar and should not be presented as a polished assistant response.

Do not represent this checkpoint as production ready, generally conversational, or safe to connect to unrestricted tools. Keep executors allow-listed and validate every typed argument before causing side effects.

Licensing and attribution

The base model is governed by the NVIDIA Open Model License Agreement, with additional upstream attribution described by the PersonaPlex model card. This repository does not redistribute the base weights. Users are responsible for reviewing and complying with the base model terms and all upstream licenses.