saggamer/KIRA-LIVE-1-AB-Experimental

KIRA Live 1 A+B — Experimental

KIRA Live 1 A+B is an experimental Apple Silicon MLX text model that combines the A agentic-planning candidate and B conversation candidate in one Thinker checkpoint. The uploaded checkpoint includes the latest nine-update conversation preference pass. It is intended for research and local testing.

This candidate did not pass the promotion gate and is not the KIRA Superapp daily driver. The full Thinker checkpoint is provided; the Listener and Talker voice weights are separate and are not included here.

Agentic performance

On the small paired evaluation, the combined parent and this updated candidate both passed 6 of 12 agentic contract-retention cases. The conversation update did not improve that score. These cases check generated text plans and selected permission fields; they do not execute tools or establish end-to-end task success.

The model can produce draft plans for web and other actions in its explicit planning mode. The included runner only prints those plans. Existing weaknesses include invented workflow anchors, unnecessary permission requests, and changed search arguments. Review plans and permissions carefully; treat this build as an experiment, not a reliable autonomous agent.

Small text check	A+B parent	Uploaded candidate
Fresh conversation cases	3/8	3/8
Multi-turn conversation retention	5/12	5/12
Agentic contract retention	6/12	6/12

The evaluation is small and uses related task families. No live audio, actual tool execution, human-EQ assessment, or comparison with Gemini was performed. The candidate showed conversation regressions despite equal aggregate counts, including repeated replies and failures on corrected memories and story constraints.

Run it

Use the model with KIRA Superapp in a separate experimental checkout. See the KIRA Live 1 production voice bundle for the current daily-driver Listener and Talker. Do not replace the production model with this experiment.

On an Apple Silicon Mac, install the provided requirements and run:

python -m pip install -r requirements.txt
python run_experimental.py --verify
python run_experimental.py --prompt "Plan a search for spooky stories."
python run_experimental.py --tree --prompt "Search for stories, but do not open any pages."

The first run obtains the pinned Qwen3.5-0.8B donor assets unless a local snapshot is provided with --donor. This custom MLX runtime is not a drop-in Transformers or MLX-LM model and does not provide a hosted inference endpoint. The --tree option prints a proposed plan only; it does not run Superapp tools.

Model and release status

The checkpoint is the full 764-tensor Thinker export with the evaluated conversation preference delta applied. Its export was checked against the trained tensor values. No additional training was performed for publication. The release is experimental, was rejected for promotion, and must not be interpreted as evidence of better overall conversation or agentic quality.

Private chats, recordings, and training examples are not included. See config.json, evaluation_summary.json, and checksums.json for provenance, measured results, and file integrity. KIRA code and included donor-derived weights retain their Apache-2.0 terms and original attribution; dependency licenses apply separately.