01 · Tileward Models

Governed · Compressed Run a bigger model in your VPC.

A model that does not fit on the card you have costs a second card, a smaller model, or a cloud bill. Tileward Models fits it.

25–78% smaller (1.3–4.5×)across the nine compressed models in the library 36% smaller (1.55×)than the publisher’s own FP8 release 5.9% smaller (1.06×)than NVIDIA’s own NVFP4 build of Qwen3.8-27B Level with FP8on a multiple-choice test, inside its error bar 3 placesit runs unchanged: hosted API, your VPC, air-gapped

Works with the clients you already run

28 have a recipe to paste, and anything else that speaks MCP or the OpenAI wire format works too.

Quality

A compressed model has three jobs. Two are measured, one is where it runs.

Fit on the hardware you have, stay the model you chose, and run wherever you run it. The first two are measured against the publisher’s own release.

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×).

Every model and its size → · the write-up

Still the model you chose

±0.3

is the error bar on each score for Tileward 35B-A3B, and the gap between our build and the publisher’s FP8 release sits inside it: level on a multiple-choice test.

The write-up → · the eval

Runs where you run it

places it runs unchanged: our hosted API, your VPC, or air-gapped.

Where each piece runs →

Estimate

What does fitting the card save?

A model that needs fewer cards costs less to run. Pick a model and the cards you have, and compare the weights as published with Tileward’s.

Estimate it 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, sizes on the catalogue below.

Setup

Provider

Copies you run

Advanced model, card, memory and rate

Model

Compare against

GPU

VRAM for context, per copyKept free for the context and the runtime, across all the cards of one copy.

Rate per card-hourUSD, yours to change. The provider fills it with its on-demand list price.

Hours per month

Line	As published	With Tileward

Model library

Every model we have compressed, in one list.

Every model we have compressed, with its size, context window, price and the id to call it by.

Tileward 35B-A3B Tileward Qwen3.8-27B Tileward PersonaPlex 7B

More are ready to serve, hosted or on your hardware. Talk to an engineer

Browse the model library →

What you get

Your hardware, your VPC, the same governance.

25% to 78% smaller (1.3× to 4.5×)

Mixtures of experts land 25% to 78% smaller than the weights they were compressed from (1.3× to 4.5×); Tileward 35B-A3B, served today, lands 66% smaller (3.0×). The dense Tileward-Qwen3.8-27b lands 63% smaller (2.7×).

OpenAI-compatible

Change the base URL and your existing client works.

Governed by the same tiles

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

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
# same call self-hosted, only the base_url changes