Geospatial runtime

mere.run geo is a local-first inference boundary for Earth-observation and humanitarian workflows. Any workflow can prepare the documented tensors, run the native model locally, and carry the resulting candidates or embeddings into its own evidence process.

Acquisition, cloud masking, reprojection, tiling, georeferencing, source citations, and analyst review remain workflow responsibilities. mere.run owns immutable model provenance, native Swift/MLX execution, hardware-aware model selection, and machine-readable outputs.

Model families

Command	Capability	Managed tiers	Automatic tier policy
geo flood	Two-class flood logits from four S2/S1/DEM observations	TerraMind Base	Base
geo fire	Two-class fire logits from four S2/S1/DEM observations	TerraMind Base	Base
geo tessera	Per-pixel Sentinel-1/2 time-series embeddings	Nano, Small, Medium, Large, 2.06B Teacher	<6 GB Nano, 6-7 GB Small, 8-11 GB Medium, 12-31 GB Large, 32+ GB Teacher
geo olmoearth	Multisensor spatial embeddings	Nano, Tiny, Small, Base	<8 GB Nano, 8-11 GB Tiny, 12-15 GB Small, 16+ GB Base

When --model is omitted, the runtime chooses the strongest tier recommended for the machine. If that tier is not installed, it selects the strongest installed tier below it. It does not silently download model weights. An explicit model ID or converted model path always overrides automatic choice.

TerraMind Fire

mere.run geo fire runs IBM and ESA's official ImpactMesh Fire checkpoint through the same audited TerraMind Base temporal segmentation graph used by the Flood checkpoint, but with independent immutable weights and provenance:

mere.run geo fire fire-input.safetensors \
  --output fire-logits.safetensors \
  --model /path/to/converted-terramind-fire \
  --json

The input must contain S2L2A, S1RTC, and DEM arrays shaped [batch, channels, 4, 256, 256], already normalized under the workflow's declared raster policy. Output logits is float32 [batch, 2, 256, 256].

TESSERA v2

mere.run geo tessera encodes raw annual Sentinel-2 and Sentinel-1 observation sequences. The four students emit Matryoshka embeddings; the full Teacher emits its native 1,024-dimensional representation:

# Hardware-selected tier and native output width.
mere.run geo tessera pixel-series.safetensors \
  --output tessera-embeddings.safetensors \
  --json

# Force the 2.06B high-memory tier.
mere.run geo tessera pixel-series.safetensors \
  --model vision-embed-tessera-v2-teacher \
  --dimensions 1024 \
  --output tessera-teacher-embeddings.safetensors

Inputs use raw values and the exact upstream band orders. Required tensors are S2 [batch, time, 10], S2_DOY [batch, time], and at least one complete S1_ASC/S1_ASC_DOY or S1_DESC/S1_DESC_DOY pair. Sentinel-1 arrays use two bands. Day of year is an unnormalized integer from 1 through 365.

Students support --dimensions 16, 32, 64, or 128. Teacher requires 1024. Teacher uses the upstream pooled normalization after merging ascending and descending Sentinel-1 observations; students use separate source-specific statistics. The runtime selects the correct contract from the immutable model variant.

The Teacher evaluates 2.064B parameters per pixel. It is valuable for focused research and distillation but is not the recommended bulk-embedding default. The Large student is the automatic ceiling below 32 GB unified memory.

OlmoEarth v1.2

mere.run geo olmoearth accepts one or more primary OlmoEarth imagery modalities and emits a spatial feature grid for each one:

mere.run geo olmoearth observations.safetensors \
  --output spatial-embeddings.safetensors \
  --patch-size 4 \
  --input-resolution 10 \
  --json

The input must contain TIMESTAMPS shaped [batch, time, 3] as (day, zero-indexed month, year), plus one or more of:

• S2L2A: [batch, height, width, time, 12]
• S1RTC: [batch, height, width, time, 2]
• LANDSAT: [batch, height, width, time, 11]

--patch-size accepts 1, 2, 4, or 8. Smaller patches retain more spatial detail and require more attention work. --include-tokens preserves the full time axis in addition to the default time-pooled grids.

OlmoEarth's artifact license permits broad environmental and humanitarian work, but prohibits military and defense applications, intelligence gathering, human surveillance and policing, and listed extractive activities. Managed pulls require --accept-model-license; the flag records review and acceptance of upstream terms, not a mere.run judgment that a proposed use qualifies.

Immutable conversion

Managed pulls download the pinned upstream checkpoint and source metadata. The Swift runtime does not interpret the Python checkpoint. Convert it once into the checksum-pinned float32 safetensors package:

python3 scripts/convert-terramind-fire-mlx.py \
  --checkpoint /path/to/TerraMind_v1_base_ImpactMesh_fire.pt \
  --configuration /path/to/terramind_v1_base_impactmesh_fire.yaml \
  --output /path/to/converted-terramind-fire

python3 scripts/convert-tessera-v2-mlx.py \
  --variant teacher \
  --checkpoint /path/to/tessera_v2_2B_teacher.pt \
  --output /path/to/converted-tessera-teacher

python3 scripts/convert-olmoearth-v12-mlx.py \
  --variant base \
  --weights /path/to/weights.pth \
  --configuration /path/to/source-config.json \
  --output /path/to/converted-olmoearth-base

Each converter rejects the wrong source hash, architecture, precision, tensor inventory, or scalar count. Each runtime loader then verifies the exact output artifact bytes and typed conversion receipt before loading weights.

For numerical verification against the official implementations, use scripts/validate-terramind-fire-reference.py, scripts/validate-tessera-v2-reference.py, and scripts/validate-olmoearth-v12-reference.py. The repository's installed-model gate constructs real input tensors and exercises the public geo commands.

Why THOR is deferred

THOR was evaluated at source commit 7ded3cea673ac21fe2bcadf9f3d9d4506eb6ab5f. Its flexible 10-1,000 m Sentinel-1/2/3 backbone is promising, especially for coarse Sentinel-3 climate and ocean observations. The evaluated release is an embedding backbone rather than a humanitarian decision head, however, and its S1/S2 role overlaps the validated TESSERA and OlmoEarth routes.

THOR is therefore deferred until a workflow needs its genuinely distinct Sentinel-3/native-resolution contract and can supply a usefulness gate for that output. Adding another generic encoder without that consumer would expand download, preprocessing, ALiBi, and validation surface without improving an implemented result.

Responsibility boundary

Flood and fire logits are candidates, not authoritative findings. Embeddings are features, not conclusions. For humanitarian promotion, retain source imagery and acquisition time, preprocessing provenance, model ID and immutable revision, output artifact hashes, corroborating evidence, and accountable human review.