> For the complete documentation index, see [llms.txt](https://docs.quantaria.xyz/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.quantaria.xyz/architecture/multimodal-backbone.md).

# Shared Multimodal Backbone

Hilbert models generate text and images from **one** set of shared tensor cores, rather than a separate text model and image model joined together.

## Unified objective

Training follows a Transfusion-style objective ([arXiv:2408.11039](https://arxiv.org/abs/2408.11039)):

* **Text** is trained with next-token prediction loss.
* **Images** are split into patches and trained with a diffusion loss.

Both losses update the same shared tensor cores, so one backbone learns to handle both modalities. In Hilbert-1, that backbone is built from MPS/MPO layers (see [Tensor-Network Weights](/architecture/tensor-network-weights.md)).

## Tensor Product Attention

Attention in Hilbert uses **Tensor Product Attention (TPA)** ([arXiv:2501.06425](https://arxiv.org/abs/2501.06425)). Queries, keys and values are factorized from the contextual activations, and the scheme is compatible with RoPE positional encoding. The paper reports a KV cache about 10x smaller. This is a result from the published paper, not a Quantaria measurement.

A smaller KV cache is what allows longer context in less memory, which is why TPA is also central to the planned Schmidt model.

## MPO mixture-of-experts

Hilbert-1 Base adds an **MPO mixture-of-experts**. Experts share one central tensor and keep small expert-specific auxiliary tensors, so adding experts costs far fewer parameters than duplicating whole feed-forward blocks. The method comes from [arXiv:2203.01104](https://arxiv.org/abs/2203.01104), which reports 27x fewer parameters than Switch Transformer. Again, this is the paper's result.

## Output sizes

|              | Hilbert-1 Nano | Hilbert-1 Base         |
| ------------ | -------------- | ---------------------- |
| Text context | 2K tokens      | 8K tokens              |
| Image output | 32×32          | 256×256                |
| Experts      | None           | MPO mixture-of-experts |

## Evaluation

Each Hilbert model is compared with a dense transformer of the same size on the same data. Hilbert-1 Nano is trained on TinyStories text and CIFAR-scale images. Metrics are loss curves, perplexity, FID and memory use, and results are published either way.

## Related

* [Models](https://github.com/bhavy2908/dataai_docs/tree/main/overview/models.md)
* [How It Works](https://github.com/bhavy2908/dataai_docs/tree/main/product/how-it-works.md)


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://docs.quantaria.xyz/architecture/multimodal-backbone.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.