saggamer/KIRA-LIVE-1

KIRA Live 1

KIRA Live 1 is a local conversational assistant for Apple Silicon Macs. It supports text and voice conversations, expressive speech, chat memory, and permission-aware actions through KIRA Superapp's THE TREE.

What you can do

• Talk naturally, pause the microphone, and interrupt spoken replies.
• Continue a conversation using its saved chat history.
• Request file inspection, web research, and app actions through THE TREE.
• Request PDF and Word documents in KIRA Superapp, with returned execution evidence rather than an unverified completion claim.

Getting started

This is a complete three-weight research-preview package for the KIRA runtime, not a drop-in Transformers model or a hosted inference endpoint. Apple Silicon and the KIRA Live Python environment are required. Use a KIRA Live-enabled Superapp checkout. Older Superapp revisions may not include the Live button; downloading this package alone does not add that UI.

1. Use KIRA Superapp.
2. Download this repository into the Superapp directory, preserving its paths.
3. Install the supplied requirements.txt. The three pinned weight sets are included under weights/; no separate donor-weight download is needed.
4. Launch RUN_KIRA_OS.command, select Talk with KIRA Live 1, tap the microphone, and wait for Listening before speaking.

The package includes the Listener, Thinker and Talker weight sets (including the speech codec), accepted KIRA checkpoints, required normalization data, runtime code and SHA-256 checksums. They run together through KIRA's native runtime, not as a single fused tensor file. This is not a claim of joint end-to-end training. Chat databases, private recordings, training examples, credentials, and rejected experiments are excluded.

For standalone voice conversation without Superapp tool execution:

python -m pip install -r requirements.txt
python run_kira_live.py --verify
python run_kira_live.py --probe
python run_kira_live.py

The first command needs network access for Python dependencies. After setup, bundled model inference is offline. Allow microphone access when prompted. Standalone mode is conversational; THE TREE tools require Superapp integration.

The current runtime includes local WebRTC acoustic echo cancellation for speaker playback, an echo-tail guard, and the multilingual token-mapping crash repair. Install the current requirements before using hands-free voice. These are runtime fixes; the accepted model weights have not been retrained.

Limitations

This preview can still make mistakes. Voice latency and interruption depend on the microphone, speakers, room noise and machine load. Tool selection, document content, permissions and long workflows need user review. No claim is made of human-level emotional intelligence or superiority over commercial Live models. Speech pipeline tests are not a guarantee of flawless hardware conversations.

License and credits

KIRA's supplied code is Apache-2.0. Third-party dependencies retain their own licenses and attribution requirements. The included donor weights retain their original authorship and Apache-2.0 terms. Exact source repositories and revisions are recorded in bundle_manifest.json; see THIRD_PARTY_NOTICES.md and LICENSE.

September 29 runtime update

Updated Live conversation history, repetition control, language selection, THE TREE action contracts, nonblocking audio events and latency diagnostics. The Talker uses in-memory 8-bit linear layers by default; set KIRA_LIVE_TALKER_BITS=0 for original BF16 precision. Saved weights are unchanged. The verified CoreML emotion package is included. See measured timings and limitations. This is a runtime update, not a new trained model or quality promotion.