README / README.md
Tileward's picture
Update Tileward-Qwen3.8-27B figures
16a38e8 verified
|
Raw History Blame Contribute Delete
1.28 kB
metadata
title: README
emoji: 🔷
colorFrom: blue
colorTo: green
sdk: static
pinned: false

Tileward

Tileward is the efficiency and control layer between applications and models. Tileward Models and Tileward Context reduce the work, by shrinking model weights and the conversation resent on every request. Tileward Documents and Tileward Governance control it, by grounding answers in your own files and declining a locked tile before the model writes a word. Run it through the hosted API, in your own VPC, or fully air-gapped.

Models

  • Tileward-Qwen3.8-27B, served on the hosted API: 20.1 GB, 64% smaller (2.8x) than Qwen3.8-27B, and level with NVIDIA's NVFP4 build of the same model on MMLU from a file 8.5% smaller.
  • Tileward-Qwen3.6-35B-A3B: 24.1 GB, 66% smaller (3.0x) than Qwen3.6-35B-A3B, and level with Qwen's own FP8 release on MMLU.

Get started

  • tileward.com: what each capability does and how it was measured.
  • app.tileward.com: create an account and an API key. Every account starts with $25 of credit and needs no card. The API is OpenAI-compatible at https://api.tileward.com/v1.