---
license: apache-2.0
base_model: prism-ml/Bonsai-8B-gguf
tags:
- 1-bit
- binary
- webgpu
- serverless
- content-addressed
- hologram
- kappa
---

# q-bonsai-8b — Bonsai-8B as a κ-object (streams into your browser like a film)

**Created using [Bonsai](https://huggingface.co/prism-ml/Bonsai-8B-gguf) by Prism ML.**

This is [prism-ml/Bonsai-8B](https://huggingface.co/prism-ml/Bonsai-8B-gguf) — a Qwen3-8B moved
end-to-end into **binary {−1,+1} weights (1.125 true bits/weight, embeddings and LM head included)** —
re-laid as a **content-addressed κ-object** for the [Hologram](https://github.com/Hologram-Technologies)
serverless substrate. **No re-quantization anywhere**: the trained sign bits pass through byte-exact
(f16 group scales widened to f32 once, exactly). The model runs **entirely in the browser on WebGPU**
— no server, no account, no data leaves the device.

## What's here

| File | What it is |
|---|---|
| `q-bonsai-8b.v1.holo` | ONE Range-streamable file: boot-ordered BLAKE3-verified blocks + embedded tokenizer. The reader serves tensors **while the download is still in flight** and warm-persists to OPFS (second visit = zero network, works offline). |
| `manifest.json` + `b/*.gz` | The same blocks as loose content-addressed parts (sha256 transport axis). |
| `manifest.blake3.json` + `sha256-to-blake3.map.json` | The canonical BLAKE3 (κ) axis. |
| `tokenizer.gguf` | The source GGUF header (tokenizer + arch), for serverless load. |

Every block is verified against its content address **before** it is decompressed or touches the GPU
(Law L5). Any static file host can serve this repo — origins are interchangeable because the bytes,
not the host, carry identity.

## Pins

- manifest κ: `did:holo:sha256:a0dc81f26ec5ce98b28ee9c1fab620e91c20720f777e1de6bc416317f54d27e2`
- root κ (tensor blocks): `sha256:8bf47176c881493a795cd911abb53603a27efc0d97ad8b5ed76f2df75b2e7816`
- `.holo` determinism witness: `sha256:7c9d809ce70ad143…` (same input dir → byte-identical file)

## Measured (first light, 2026-07-15)

Chrome / WebGPU, consumer laptop GPU: engine resident **7.8 s** after open (progressive reader
returned at **1%** of the wire); decode **14–25 tok/s**; warm second visit loads from OPFS with
**zero network**. Coherent Qwen3 thinking-mode output.

## Provenance

Weights: Prism ML's Bonsai-8B (Apache-2.0) — see `LICENSE` and `NOTICE.txt`.
Conversion: `compile2bit.mjs q1` mode (pass-through re-layout) + `holo-kappa-pack.mjs`, part of the
Hologram Q substrate. Format `q1`: blob = `[signs N·K/8 B][f32 scales N·K/128·4 B]` per tensor;
GEMV kernel dequantizes inside the matmul (weights never expand to dense f32 in memory).