# TESSERA v1 model weights now available

Date: 2026-03-10
Author: Zhengpeng Feng
Canonical URL: https://geotessera.org/blog/2026-03-10-tessera-v1-weights

# TESSERA v1 model weights now available

We are pleased to announce that the TESSERA v1 model weights are now publicly available for download. Both the full-precision and quantized checkpoints can be used to generate your own embeddings from Sentinel-1 and Sentinel-2 imagery.

## What's available

**Full-precision checkpoint** — generates float32 embeddings, suitable for research and experimentation.

**QAT checkpoint** — trained with quantization-aware training, outputs int8 embeddings with per-pixel float32 scales. This is the same model used to produce the precomputed global embeddings in the [geotessera](https://pypi.org/project/geotessera/) library. The QAT pipeline supports both CPU and GPU inference, with automatic AMX acceleration on supported hardware.

Download both checkpoints from the [TESSERA v1 release on GitHub](https://github.com/ucam-eo/tessera/tree/beta_version_1.0#qat-model-weight-quantized-output).

## Quantized output format

The QAT model produces two files per tile:

- `tile_name.npy` — int8 embedding tensor, shape `(H, W, 128)`
- `tile_name_scales.npy` — float32 scale map, shape `(H, W)`

This compact format reduces storage and memory requirements while preserving embedding quality for downstream tasks.

## License

All TESSERA model weights and embeddings are released under [CC0](https://creativecommons.org/publicdomain/zero/1.0/) — free for both commercial and non-commercial use. The software is MIT-licensed. While attribution is not legally required, we appreciate it.

## Getting started

See the full [inference pipeline documentation](https://github.com/ucam-eo/tessera/tree/beta_version_1.0) for setup instructions, or use the [geotessera](https://geotessera.readthedocs.io) Python library to work with precomputed embeddings directly.