---
license: mit
base_model: microsoft/bitnet-b1.58-2B-4T
pipeline_tag: text-generation
inference: false
tags:
- hologram
- q
- kappa-object
- holo
- ternary
- bitnet
- llama3
---

<div align="center">

# Hologram · BitNet-2B-4T

**Native 1.58-bit ternary brain**

`t2 · 1.58-bit ternary` · `0.69 GB` · streamed to Q as a **key-addressable `.holo` object**

[Hologram](https://gethologram.ai) · [Live Space](https://huggingface.co/spaces/HOLOGRAMTECH/hologram) · [Organization](https://huggingface.co/HOLOGRAMTECH) · [Code](https://github.com/Hologram-Technologies)

</div>

---

## What this is

Microsoft's BitNet b1.58 2B4T, the first natively 1.58-bit trained model at scale, re-encoded to Hologram's ternary key format. Q's default brain: fast, tiny, and coherent.

This repository is **not** a GGUF or Transformers checkpoint. It is a **Hologram key object**: the weights of `microsoft/bitnet-b1.58-2B-4T` re-encoded into Hologram's content-addressed `.holo` format so they stream, one verified block at a time, into **Q**, the on-device brain of the Hologram web OS. It runs in the browser on WebGPU, serverless, with nothing to install.

## How it streams

The object is laid out for cold streaming from an untrusted CDN:

| File | Role |
|---|---|
| `manifest.json` | the root. Names every tensor and the key (content hash) of its block. |
| `b/sha256_*.gz` | the tensor blocks. Each filename **is** the SHA-256 of its bytes. |
| `tokenizer.gguf` | bundled header, so loading is fully serverless. |

Q fetches the manifest, then pulls each block by its key and re-derives `sha256(block)` on arrival. If a byte is wrong, the block is rejected. Nothing is trusted; everything is proven.

## Verify (Law L5)

The object's identity is the SHA-256 of its manifest, pinned in Q's catalog before a single byte of weight is trusted:

```
did:holo:sha256:fcf835659d88d2fe6f683cf1ab8de6a6ba6214ea0deeee4b1bcf3da1a4c05412
```

```bash
curl -sL https://huggingface.co/HOLOGRAMTECH/q-bitnet-2b/resolve/main/manifest.json | sha256sum
```

## Specifications

| | |
|---|---|
| Architecture | BitNet b1.58 (Llama 3 template) |
| Precision | t2 · 1.58-bit ternary |
| Object size | 0.69 GB |
| Hidden size | 2560 |
| Layers | 30 |
| Heads (Q / KV) | 20 / 5 (GQA) |
| FFN | 6912 |
| Vocab | 128256 |
| Context | 3000 |
| Format | `holo-2bit/1` |

## Provenance and license

Derived from [`microsoft/bitnet-b1.58-2B-4T`](https://huggingface.co/microsoft/bitnet-b1.58-2B-4T). Inherits the MIT license from microsoft/bitnet-b1.58-2B-4T. The re-encoding is content-addressed at the key level: the object either re-derives to its pinned identity or it is refused.

## Run it

These weights load through Q, not a standard runtime. Open the [Live Space](https://huggingface.co/spaces/HOLOGRAMTECH/hologram) or visit [gethologram.ai](https://gethologram.ai) to run Hologram, then pick **BitNet-2B-4T** from Q's model list.

<div align="center"><sub>Composed on the golden ratio. One key, everything.</sub></div>