Allen Institute Releases OlmoEarth v1.1 with 3x Compute Reduction for Satellite Imagery
Allen Institute (AI2) released OlmoEarth v1.1, a family of transformer-based models for satellite imagery processing that reduces compute costs by up to 3x compared to the original OlmoEarth v1. The efficiency gains come from collapsing Sentinel-2 resolution bands into single tokens, cutting sequence lengths by three times while maintaining benchmark performance.
OlmoEarth v1.1: 3x Compute Reduction for Satellite Imagery Models
Allen Institute for AI (AI2) released OlmoEarth v1.1 on May 19, 2026, a family of transformer-based models that cuts compute costs by up to 3x compared to OlmoEarth v1 while maintaining performance on remote sensing benchmarks.
Technical Implementation
The efficiency gains stem from a fundamental redesign of how the model tokenizes Sentinel-2 satellite imagery. OlmoEarth v1 created separate tokens for each of Sentinel-2's three resolutions (10m, 20m, and 60m), generating 6 tokens per spatial patch per 2-timestep input. OlmoEarth v1.1 collapses these into single tokens, reducing token counts by three times.
This approach wasn't straightforward. AI2 reports that naive token merging caused a 10 percentage point drop on m-eurosat kNN, a standard remote sensing benchmark. The team modified their pretraining regimen to maintain cross-band relationship modeling, detailed in their technical report.
Model Family
AI2 released three model sizes:
- Base
- Tiny
- Nano
All variants process Sentinel-2 data with tensors formatted as [H, W, T, D=12], where H and W represent latitudinal and longitudinal pixels, T is the temporal dimension, and D covers 12 Sentinel-2 channels.
Performance Trade-offs
AI2 states OlmoEarth v1.1 maintains similar performance to v1 on their benchmark mix and partner-constructed tasks, though the technical report notes "some regressions." Because both versions train on identical datasets, performance differences isolate methodological changes.
Compute costs in transformer models scale quadratically with token sequence length, making the 3x reduction in tokens significant for inference and fine-tuning. AI2 measured efficiency using MACs (multiply-accumulate operations per forward pass).
Deployment Context
Since OlmoEarth v1's November 2025 release, partners have deployed it for mangrove tracking, forest loss classification, and country-scale crop mapping. AI2 reports deployments now scale to national, continental, and global areas, with data export, preprocessing, inference, and post-processing dominated by compute costs.
What This Means
OlmoEarth v1.1 addresses the practical bottleneck in satellite imagery AI: compute cost at scale. A 3x reduction enables more frequent planet-scale map refreshes and lowers barriers for organizations without large compute budgets. For researchers, training on identical datasets to v1 creates a controlled comparison for studying pretraining methodologies in remote sensing.
The model family is available on Hugging Face with full training code. AI2 recommends existing OlmoEarth v1 users test v1.1 for their specific tasks given the documented performance regressions.
Related Articles
Ai2 open-sources AstaBrief 8B, a Qwen3-8B report model it says runs 3.5x faster than Claude in Asta
Ai2 has open-sourced AstaBrief 8B, a model fine-tuned from Qwen3-8B that turns a research question and retrieved literature excerpts into a cited report. It is live in Asta as Fast mode, which averages 51.1 seconds per report versus 178.5 seconds for the Claude-powered Thinking mode, according to Ai2. The weights and training data are public.
China Telecom's Xing4.0-29B-A4B: 29B MoE, 4B Active, 256K Context, Trained Fully on Ascend NPUs
China Telecom AI's Xing4.0-29B-A4B (formerly the TeleChat line) is a mixture-of-experts model with 29B total and 4B active parameters and a native 256K context window, extensible to 512K. The company claims it is the first model of this scale trained entirely on Ascend NPUs with MindSpore. Community GGUF quantizations from Venastine-Research are already available.
Cloudflare releases Clef decision models, claims 39 ms median latency vs. 524 ms for TypeSafe's Jev
Cloudflare has released Clef and Clef-flash, two open-weight decision models that return probabilities over predefined answer options instead of generating text. The company claims median latencies of 39 ms and 209 ms, against just over 524 ms for TypeSafe AI's Jev. Both support text and images and are API-compatible with Jev.
inclusionAI releases Ling 3.1 Flash: 560B MoE, 25B active, 262K context, free on OpenRouter
inclusionAI has released Ling 3.1 Flash, a hybrid reasoning mixture-of-experts model with 560B total and 25B active parameters and a 262K-token context window. It is listed as free on OpenRouter through NovitaAI. No benchmark scores have been published on the listing.
Comments
Loading...