# Tileward Models: run a bigger model in your VPC, on an air-gapped network, or on the GPUs you have

Two thirds less memory, wherever you run it. A model that does not fit on the card you have costs a second card, a smaller model,
or a cloud bill; Tileward Models compresses it so it fits, and it runs the same on the hosted API, in your VPC or air-gapped.

Page: https://tileward.com/models/

## What compression does

- **Fits the card you have**: 25–78% smaller (1.3–4.5×) than the weights as published, across the nine models in the library that we have compressed. A mixture of experts lands 25% to 78% smaller (1.3× to 4.5×); a dense model lands 63% smaller (2.7×).
- **Still the model you chose**: Tileward 35B-A3B is 36% smaller (1.55×) than the publisher's own FP8 release, and level with it on a multiple-choice test: the gap sits inside the ±0.3 error bar on each score. [The eval](https://tileward.com/evals/#compressed-model), [the write-up](https://tileward.com/research/what-compression-inherits/).
- **Smaller than NVIDIA's own build**: Tileward-Qwen3.8-27b is 5.9% smaller than NVIDIA's NVFP4 build of Qwen3.8-27B (1.06×), and level with it on MMLU. [The eval](https://tileward.com/evals/#compressed-model-27b).
- **Runs where you run it**: 3 places, unchanged: the hosted API, your VPC, or air-gapped. [Where each piece runs](https://tileward.com/platform/).

## Served today

- Tileward 35B-A3B, served today, is 66% smaller (3.0x) than the weights it was compressed from; the dense Tileward-Qwen3.8-27b is 63% smaller (2.7x).
- Every model, its size, context window, price and id: [the model library](https://tileward.com/model-library/) ([markdown](https://tileward.com/model-library/index.md)).
- More are ready to serve, hosted or on your hardware: [talk to an engineer](https://calendar.app.google/nt8a67yTAWrbB44o6).

## How you call it

OpenAI-compatible: change the base URL and your existing client works. Self-hosted, only the `base_url` changes.

```python
client = openai.OpenAI(base_url="https://api.tileward.com/v1", api_key=TILEWARD_KEY)
client.chat.completions.create(model="Tileward-Qwen3.6-35B-A3B", messages=[...])
# also answers to tileward-35b-a3b and Qwen3.6-35B-A3B-TW, at the same rate
```

The 215 Tileward Governance tiles apply to the self-hosted model exactly as they do to the hosted API, with audit records retained locally.

## Three ways to run it

| | Hosted API | Your VPC | On-prem, air-gapped |
|---|---|---|---|
| Time to run | Minutes | Days | Weeks |
| What it takes | Change the base URL | Single node, one GPU | Runs on your hardware, fully offline |
| Governance | Default tile pack | Tiles and policies of your own | Local audit retention |

## Estimate the saving

The calculator on the page compares the cards a model needs as published with the cards it needs from Tileward. By hand: divide the
model's weights by what one card holds after the VRAM you set aside for context, round up, and multiply by the copies you run and
what a card costs an hour; do it once with the weights as published and once with Tileward's. [Calculator](https://tileward.com/models/#calculator).

## Related

- [Start free, no card](https://app.tileward.com/?join=explore)
- [Talk to an engineer about your hardware](https://calendar.app.google/hg2DzFTxKJgwFWm19)
- [Pricing](https://tileward.com/pricing/)