TESSERA v2 — Large pixel student (43.83 M)

Name: TESSERA-V-2.0-2B-L
Availability: InStock

A compact per-pixel encoder for Sentinel-2 + Sentinel-1 satellite time series, distilled from the TESSERA v2 2B teacher. It maps one pixel's annual observation history to a 128-dimensional Matryoshka embedding: the first K dimensions are independently usable for K ∈ {16, 32, 64, 128}, so you can truncate to 16/32/64 dims with no retraining and no separate checkpoint.

Part of the TESSERA v2 release described in TESSERA v2: Scaling Pixel-wise Earth Foundation Models.

Where this model comes from

The teacher — geotessera/TESSERA-V-2.0-2B-Teacher — is a 2,064,266,242-parameter pixel-wise encoder producing 1024-d representations. Evaluating 2.06 B parameters per pixel makes it impractical to deploy at tile or global scale, so it is released mainly as a distillation target. This student is one of four compact encoders distilled from it for actual use.

Distillation uses a Matryoshka objective: a separate linear head per prefix length K ∈ {16, 32, 64, 128} reconstructs the full frozen teacher embedding from the student's first K coordinates. That is what imposes an ordering on the student's coordinates — a redundancy-reduction objective like Barlow Twins identifies subspaces only up to a rotation, so self-supervised prefix losses alone cannot produce a usable nesting. The projection heads are training-only and are discarded at inference; this checkpoint contains the encoder alone.

Note on normalization. The students z-score Sentinel-1 ascending and descending with their own per-source statistics before merging the two into one stream. The teacher instead merges them in raw units and applies a single set of pooled statistics. If you move data between the two, do not carry the normalization across.

The family

Variant	Repo	Parameters	`latent_dim`	`d_model`	Layers	FFN
Nano	`geotessera/TESSERA-V-2.0-2B-N`	1.07 M	36	144	2	384
Small	`geotessera/TESSERA-V-2.0-2B-S`	7.11 M	64	256	4	1024
Medium	`geotessera/TESSERA-V-2.0-2B-M`	21.03 M	110	440	4	1792
Large	`geotessera/TESSERA-V-2.0-2B-L`	43.83 M	160	640	4	2560

All four share identical inference code and the same 128-d Matryoshka output contract — only capacity differs. Nano targets edge/on-device use, Medium is the balanced default, Large is intended for provider-side global inference.

Architecture

`latent_dim`	`d_model`	layers	heads	FFN	output	parameters
160	640	4	4	2560	128 (Matryoshka)	43,831,170

Two per-modality backbones (Sentinel-2, and ascending+descending Sentinel-1 merged into one stream). Each backbone is an MLP band embedding plus a sinusoidal day-of-year positional encoding, a post-LN Transformer encoder (ReLU FFN, QK-norm off), and single-head softmax attention pooling over time. The two pooled vectors are concatenated and passed through an MLP dim_reducer that ends in a non-affine LayerNorm — so every output embedding is exactly mean 0 / std 1 across its 128 dimensions.

Input contract

[!IMPORTANT] The Sentinel-2 channel order is not the conventional ascending-wavelength order. Feeding bands in the usual B02→B12 order will silently produce garbage embeddings. The model expects exactly:

B04 B02 B03 B08 B8A B05 B06 B07 B11 B12

Sentinel-1 is VV VH, ascending and descending concatenated along time.

Sentinel-2: (T, 10) per pixel, raw L2A reflectance (unscaled DN), plus a (T,) day-of-year array and a (T,) validity mask (1 = clear, 0 = cloud).
Sentinel-1: (T, 2) per pixel, raw RTC values, plus day-of-year. All-zero timesteps are treated as missing.
Day-of-year is passed as a raw integer 1–365 (not normalized).

infer.py performs standardization internally (standardize=True, the default): Sentinel-2 uses per-band statistics, and Sentinel-1 ascending and descending each use their own statistics, applied before the two are merged into a single stream. Pass standardize=False only if your arrays are already z-scored with those exact constants (they are in model.py).

Per pixel, the number of valid observations is bucketized to the nearest bin in {8, 16, 24, ..., 256}, then padded or subsampled to that bin size — matching the training-time procedure. Pixels sharing a bin are batched together.

Usage

python

from huggingface_hub import snapshot_download
import sys, torch, numpy as np

path = snapshot_download("geotessera/TESSERA-V-2.0-2B-L")
sys.path.insert(0, path)

from model import load_model
from infer import encode_tile, encode_pixels
from quantize import quantize as quantize_int8

model = load_model(f"{path}/ckpt/student_large.pt", torch.device("cuda"))

# One tile -> (H, W, 128) embedding map.
#   s2_bands (T,H,W,10) raw reflectance, s2_doys (T,), s2_masks (T,H,W) 1=valid
#   s1_asc / s1_desc  (T,H,W,2) raw + their own doys
emb = encode_tile(
    model, s2_bands, s2_doys, s2_masks=s2_masks,
    s1_asc_bands=s1_asc, s1_asc_doys=s1_asc_doys,
    s1_desc_bands=s1_desc, s1_desc_doys=s1_desc_doys,
    batch_pixels=4096, device=torch.device("cuda"),
)                                    # -> (H, W, 128) float32

emb16 = emb[..., :16]                # Matryoshka truncation, 1/8 the storage

code, scale = quantize_int8(emb)     # int8 (H,W,128) + per-pixel float32 scale
recon = code.astype("float32") * scale[..., None]

For per-pixel time series rather than a gridded tile, call encode_pixels() on (N, T, C) arrays — same keyword arguments, returns (N, 128).

Files

TESSERA-V-2.0-2B-L/
├── model.py          PixelStudent + load_model()  (torch + numpy only)
├── infer.py          encode_pixels() / encode_tile()
├── quantize.py       linear int8 quantize() / dequantize()
├── __init__.py       package-style re-exports
├── requirements.txt  torch>=2.0, numpy>=1.21
└── ckpt/student_large.pt

The checkpoint is a plain torch.save payload with two keys, model (the state_dict) and args (the architecture config load_model reads). No custom classes are pickled, and nothing outside torch and numpy is needed.

Intended use and limitations

Intended for scientific research and downstream geospatial analysis: environmental monitoring, conservation and habitat mapping, biomass and carbon estimation, crop mapping and agricultural monitoring, land-use and land-cover work.

Limitations:

Embeddings are annual, 10 m spectral-temporal representations — not raw imagery, and not a real-time monitoring product.
Quality degrades where a pixel has very few valid observations (persistent cloud, sparse revisit).
Validate on your own task and geography before operational use.
No benchmark scores are quoted here. The results in the TESSERA v2 preprint were measured on students distilled from the 1B teacher; these checkpoints are distilled from the 2B teacher and their evaluation will be published separately.

License

Released under CC0 1.0 (public domain dedication), matching the rest of the TESSERA release. Use of TESSERA is additionally governed by the project's Acceptable Use Policy.

Citation

bibtex

@article{tessera_v2_2026,
  title   = {TESSERA v2: Scaling Pixel-wise Earth Foundation Models},
  author  = {Feng, Zhengpeng and Jaffer, Sadiq and Shokar, Ira and
             Knezevic, Jovana and Elvers, Mark and Atzberger, Clement and
             Young, Robin and Naik, Aneesh and Robinson, Niall and
             Blake, Andrew and Coomes, David and Madhavapeddy, Anil and
             Keshav, Srinivasan},
  journal = {arXiv preprint arXiv:2607.03949},
  year    = {2026},
  url     = {https://arxiv.org/abs/2607.03949}
}

geotessera/
TESSERA-V-2.0-2B-L