q-bonsai-8b-speaking

The conversing mind — one κ

One hash pins an intelligence you can talk to:

blake3 root = 4eed444d6f04bcfc059f66bab17b7a54d2eca96b7a8713f8f5eeab0b3d14be99

payload.bin (1.30 GB, 428 members) is Bonsai-8B (Q1_0, 1-bit), the pocket-tts voice, and the moonshine ear — model weights, VAD, tokenizers, and the ear's entire pinned runtime — concatenated as content-addressed blocks in experience order: the voice first (resident at 6% of the wire, so it narrates the rest of its own arrival), then the ear, then the mind. payload.obao (1.2 MB, 0.098%) is the BLAKE3 outboard: a browser verifies every 64 KiB of any HTTP Range read against the one root before a byte is trusted. index.json is the carve map — including per-device slices (only: mobile/desktop): one root pins every device's slice; a phone streams moonshine-tiny, a desktop moonshine-base, nobody downloads what they don't run.

Talk to it: https://hologramtech-bonsai-kappa-stream.static.hf.space — tap the mic; your speech is segmented (Silero VAD), transcribed on-device (~300 ms, moonshine), answered by the 8B, and spoken back clause by clause while it writes. Audio never leaves the browser. Repeat visits are 0-network and work offline.

Provenance

• Bonsai-8B: PrismML (Apache-2.0) → κ-object HOLOGRAMTECH/q-bonsai-8b.
• TTS by Kyutai: kyutai/pocket-tts (code MIT, weights CC-BY-4.0, voice alba); int8 ONNX export KevinAHM/pocket-tts-onnx (Apache-2.0) → HOLOGRAMTECH/q-pocket-tts.
• STT: Moonshine © Useful Sensors (MIT), ONNX export onnx-community/moonshine-base-ONNX / moonshine-tiny-ONNX (q8). VAD: Silero VAD (MIT), onnx-community export.
• Runtime: @huggingface/transformers 3.8.1 (Apache-2.0) + its onnxruntime-web wasm, pinned inside the object so the ear runs with zero CDN and offline.
• Minter: holo-bonsai/bao/mint-fused.mjs (hbao — ~150 lines on blake3::hazmat, 64 KiB groups; root == plain blake3(payload)).