Skip to main content

Lab results · NVIDIA

How NVIDIA's Nemotron models do on geospatial work

We benchmarked six NVIDIA Nemotron models on GIS and Earth observation work. Here is where they are strong, where they fall behind, and what they cost.

Research

NVIDIA's Nemotron 3 Ultra scores 69 out of 100 on Earth-Bench V2, our benchmark of AI agents analysing satellite and Earth observation data. DeepSeek V4 Pro and GLM-5.2, at the same high reasoning level in the same harness, score 68. Nemotron 3 Ultra gets there at about a fifth of DeepSeek V4 Pro's cost per task.

The harness is Terminus-2, Terminal-Bench's open harness, our default. Claude Opus 4.7 at high reasoning, in Anthropic's own coding agent Claude Code, also scores 69, at about five times Nemotron 3 Ultra's cost per task. Results as of 16 November 2026.

On Earth observation data, Nemotron 3 Ultra is level with DeepSeek V4 Pro

  1. Kimi K3 (Reasoning)Terminus-275
  2. DeepSeek V4 Flash (high)Terminus-273
  3. Nemotron 3 Ultra (high)Terminus-269
  4. Claude Opus 4.7 (high)Claude Code69
  5. DeepSeek V4 Pro (high)Terminus-268
  6. GLM-5.2 (high)Terminus-268
  7. Qwen 3.6 35B-A3B (Reasoning)Terminus-262
  8. Muse Glimmer 30B (high)Terminus-260
  9. Step 5 Preview (Reasoning)Terminus-260
  10. Mistral Large 4 (high)Terminus-250
  11. Nemotron 3.5 Super 120B (Reasoning)Terminus-249
  12. Gemma 4 26B A4B (Reasoning)Terminus-246
  13. Nemotron 3.5 Super VL 120B (Reasoning)Terminus-246
  14. Nemotron 3 Nano Omni 30B (Reasoning)Terminus-231
  15. Nemotron 3.5 Lightning 30B (Reasoning)Terminus-216
  16. Nemotron 3.5 Bolt E4B (Reasoning)Terminus-25
Earth-Bench V2 score, every NVIDIA model against open weights models in the same harness and Claude Opus 4.7 in Claude Code. Results as of 16 November 2026. Source: Earth-Bench V2.

Which models we ran, and on what

We benchmark six NVIDIA models, all open weights: Nemotron 3 Ultra, Nemotron 3.5 Super 120B, Nemotron 3.5 Super VL 120B, Nemotron 3 Nano Omni 30B, Nemotron 3.5 Lightning 30B and Nemotron 3.5 Bolt E4B. NVIDIA has no coding agent of its own, so every Nemotron entry runs in Terminus-2. Each entry gets one scored attempt per task, in a sandbox with no internet.

The headline score is the Geospatial Intelligence Index, a score out of 100 across three groups of work:

  • Agent tasks (4 of 6 parts): Earth-Bench V2 (satellite and Earth observation data), GeoBenchX (GIS tasks that end in a map, a number or a list) and GS-QA2 (questions about places and terrain, answered by querying a database of OpenStreetMap features and elevation).
  • Guided workflows (1 of 6): GeoAgentBench and GeoAnalystBench, GIS workflows where the task gives the steps and the model writes and runs the code.
  • Geospatial knowledge (1 of 6): questions about GIS and Earth observation, answered with no tools.

41 models are ranked on the index. Nemotron 3 Ultra scores 73.7, the highest of the NVIDIA models.

Geospatial Intelligence Index, Text only (0 to 100, higher is better)

  1. Claude Opus 4.7 (high)80.8
  2. DeepSeek V4 Pro (max)79.7
  3. GLM-5.2 (high)78.6
  4. Qwen 3.8 27B (xhigh)75.8
  5. Nemotron 3 Ultra (high)73.7
  6. Mistral Large 4 (high)71.3
  7. Nemotron 3.5 Super 120B (Reasoning)66.3
  8. Gemma 4 26B A4B (Reasoning)64.4
  9. Nemotron 3.5 Super VL 120B (Reasoning)59.0
  10. gpt-oss-120b (high)51.9
  11. Nemotron 3 Nano Omni 30B (Reasoning)31.9
  12. Nemotron 3.5 Lightning 30B (Reasoning)27.4
  13. Nemotron 3.5 Bolt E4B (Reasoning)16.5
Every NVIDIA model, with the best model from other labs near it. Each model at its best ranked level. Geospatial Intelligence Index. Results as of 16 November 2026. Source: Geospatial Intelligence Index.

Where Nemotron 3 Ultra is strong

Earth observation data. On Earth-Bench V2 the agent works through satellite and Earth observation data: vegetation indices, land surface temperature, land cover. Among open weights models in Terminus-2, only Kimi K3 (75), DeepSeek V4 Flash (73) and Kimi K2.7 Code (70) score higher than Nemotron 3 Ultra's 69. Gaps under about 3 points are within our noise, so it is level with DeepSeek V4 Pro, GLM-5.2 and Claude Opus 4.7 here.

Knowing when a task cannot be done. GeoBenchX mixes solvable tasks with tasks that cannot be done with the data given, and a correct answer to those is to say so. On those impossible tasks Nemotron 3 Ultra scores 68, against 62 for DeepSeek V4 Pro and 53 for GLM-5.2, both at high reasoning in the same harness. It answers 32% of them anyway, against 37% and 45%. On the solvable tasks it is behind (62, against 68 and 76), so its total of 65 lands level with DeepSeek V4 Pro.

Following a given workflow. On GeoAgentBench, where the steps are spelled out, it scores 96, the same as DeepSeek V4 Pro and Claude Opus 4.7.

Nemotron 3 Ultra against DeepSeek V4 Pro, both at high reasoning in Terminus-2 (0 to 100)

Nemotron 3 Ultra (high)DeepSeek V4 Pro (high)

  1. Earth-Bench V26968
  2. GeoBenchX6565
  3. GS-QA27077
  4. GeoAgentBench9696
  5. GeoAnalystBench63100
The three agent benchmarks, then the two guided workflow benchmarks. Same harness, same reasoning level. Results as of 16 November 2026. Source: Nemotron 3 Ultra (high), every benchmark.

Where it falls behind

Spatial questions over a database. On GS-QA2 Nemotron 3 Ultra scores 70. DeepSeek V4 Pro scores 77 and GLM-5.2 78 at the same level and harness. The agent answers questions such as the nearest park to a landmark or the slope at a point by querying a PostGIS database of OpenStreetMap features and elevation. We have not yet traced which step it misses most.

Longer analysis workflows. On GeoAnalystBench it scores 63, where DeepSeek V4 Pro and GLM-5.2 both score 100. That pulls its guided workflow score down to 80, against 98 for DeepSeek V4 Pro at max reasoning.

The smaller Nemotron models

The smaller models know their geography. On geospatial knowledge, Nemotron 3.5 Super 120B scores 93, the same as Claude Opus 4.7; the best score of any model is 95. Super VL 120B scores 92, Lightning 30B 90 and Nano Omni 30B 85.

Multi-step agent work separates them. On agent tasks Super 120B scores 56 and Super VL 120B 52. Nano Omni 30B scores 21, Lightning 30B 13 and Bolt E4B 4. On GS-QA2, Lightning 30B scores 4 while it scores 90 on knowledge. An agent task can run up to 60 turns of reading data, writing code and checking the output; knowing the answer to a question is a different skill. We have not yet traced where in those turns the smaller models go wrong.

Super 120B is the one to look at for cost. It scores 66.3 on the index, 7 points below Nemotron 3 Ultra, at under a fifth of its cost per agent task. It also scores 81 on guided workflows, ahead of Nemotron 3 Ultra's 80.

What it costs

Cost per task is what one attempt would cost at the provider's list prices, failed attempts included. On Earth-Bench V2, Nemotron 3 Ultra costs about a fifth of DeepSeek V4 Pro (high) and of Claude Opus 4.7 (high) for the same score, and under a third of GLM-5.2 (high).

Across all agent tasks it costs under a quarter of Claude Opus 4.7 per task. Claude Opus 4.7 scores 7 points more on the index (80.8 against 73.7), so the trade is real: you pay more for a stronger GIS and database model, and about the same for Earth observation.

All six Nemotron models are open weights, so you can also run them on your own servers. Our costs use the list prices of hosted providers; on your own hardware the cost depends on that hardware.

What this means if you are choosing a model

  • Satellite and Earth observation analysis on an open model: Nemotron 3 Ultra is level with the best open models in the same harness, at a fraction of their cost per task. Put it on your shortlist.
  • GIS work over vector data and databases: DeepSeek V4 Pro and GLM-5.2 score higher on GS-QA2 and GeoAnalystBench. Compare them on your own tasks.
  • A small model on limited hardware: Nemotron 3.5 Super 120B keeps most of Nemotron 3 Ultra's score at a fraction of the cost. The 30B and smaller models know the field but do not yet complete multi-step agent tasks reliably.

Every score for every model is on the Geospatial Intelligence Index. If you want to know how these models do on a task your team runs today, send it to us and we will benchmark it.