Skip to main content

Lab results · MiniMax

How MiniMax's M3 models do on geospatial work

MiniMax M3.1 Flash is 7 points ahead of gpt-oss-120b at two thirds of the cost. MiniMax M3, in its own coding agent, is level with Claude Sonnet 5 on Earth observation data.

Research

MiniMax M3.1 Flash scores 53 out of 100 on the Geospatial Agent Index, the average of our three geospatial agent benchmarks. OpenAI's gpt-oss-120b at high reasoning, in the same harness, scores 46. MiniMax M3.1 Flash is 7 points ahead at about two thirds of the cost per task.

The harness is Terminus-2, Terminal-Bench's open harness, our default. Both models are open weights. Results as of 16 November 2026.

MiniMax M3.1 Flash is 7 points ahead of gpt-oss-120b in the same harness

  1. Kimi K3 (Reasoning)Terminus-273
  2. DeepSeek V4 Flash (high)Terminus-272
  3. GLM-5.3 (max)Terminus-272
  4. MiniMax M3 (Reasoning)MiniMax Code68
  5. Nemotron 3 Ultra (high)Terminus-268
  6. Qwen 3.8 27B (xhigh)Terminus-267
  7. Muse Glimmer 30B (high)Terminus-264
  8. Mistral Large 4 (high)Terminus-263
  9. Gemma 4 26B A4B (Reasoning)Terminus-256
  10. MiniMax M3.1 Flash (max)Terminus-253
  11. Step 5 Preview (Reasoning)Terminus-252
  12. gpt-oss-120b (high)Terminus-246
Geospatial Agent Index, each open weights model's best entry in Terminus-2, plus MiniMax M3 in MiniMax Code. Results as of 16 November 2026. Source: Geospatial Agent Index.

Which models we ran, and on what

We benchmark two MiniMax models, both open weights:

  • MiniMax M3.1 Flash at max reasoning, in Terminus-2.
  • MiniMax M3 with reasoning on, in MiniMax's own coding agent, MiniMax Code. A run of M3 in Terminus-2 is still being benchmarked; its results are provisional and not ranked, so this article leaves them out.

Each entry gets one scored attempt per task, in a sandbox with no internet. The three agent benchmarks are 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).

The Geospatial Intelligence Index adds guided workflows (GeoAgentBench and GeoAnalystBench, where the task gives the steps) and geospatial knowledge, weighted 4 to 1 to 1. 41 models are ranked on it. MiniMax M3.1 Flash scores 61.3. MiniMax M3 is not on the index yet.

Where M3.1 Flash is strong

Spatial questions over a database, at low cost. On GS-QA2 the agent answers questions such as the nearest park to a landmark or the slope at a point, by writing queries against a PostGIS database. M3.1 Flash scores 71, against 42 for gpt-oss-120b. Nemotron 3 Ultra and Muse Glimmer 30B score 70 in the same harness, and M3.1 Flash costs about a fifth of Nemotron 3 Ultra per task and under a fifth of Muse Glimmer 30B.

Following a given workflow. On GeoAgentBench, where the steps are spelled out, it scores 88, against 70 for gpt-oss-120b.

Not answering what cannot be answered. GeoBenchX mixes solvable tasks with tasks that cannot be done with the data given, and a correct answer to those is to say so. M3.1 Flash answers 12% of those impossible tasks anyway, against 33% for gpt-oss-120b. Its GeoBenchX total is level, 46 against 45.

MiniMax M3.1 Flash against gpt-oss-120b, both in Terminus-2 (0 to 100)

MiniMax M3.1 Flash (max)gpt-oss-120b (high)

  1. Earth-Bench V24150
  2. GeoBenchX4645
  3. GS-QA27142
  4. GeoAgentBench8870
  5. GeoAnalystBench4753
The three agent benchmarks, then the two guided workflow benchmarks. Same harness. Results as of 16 November 2026. Source: MiniMax M3.1 Flash (max), every benchmark.

Where it falls behind

Earth observation data. On Earth-Bench V2 it scores 41, against 50 for gpt-oss-120b. Earth-Bench V2 has the agent work through satellite rasters: vegetation indices, land surface temperature, land cover.

Longer analysis workflows. On GeoAnalystBench it scores 47, against 53 for gpt-oss-120b. That pulls its guided workflow score down to 68, against 87 for Step 5 Preview and 86 for Qwen 3.8 27B and Mistral Large 4.

Against the stronger open models. Kimi K3, DeepSeek V4 Flash and GLM-5.3 score 72 to 73 on the Agent Index in Terminus-2, about 20 points more than M3.1 Flash. They also cost about 4 to 21 times as much per task.

MiniMax M3 in MiniMax Code

In its own coding agent, MiniMax M3 scores 68 on Earth-Bench V2. Claude Sonnet 5 at high reasoning in Claude Code also scores 68, and Claude Opus 4.7 at high scores 69. Gaps under about 3 points are within our noise.

Earth-Bench V2 in the labs' own coding agents (0 to 100, higher is better)

  1. Claude Opus 5.5 (max)Claude Code80
  2. Kimi K3 (Reasoning)Kimi Code75
  3. GPT-6 Sol (high)Codex75
  4. Qwen 3.8 27B (Reasoning)Qwen Code69
  5. Claude Opus 4.7 (high)Claude Code69
  6. MiniMax M3 (Reasoning)MiniMax Code68
  7. Claude Sonnet 5 (high)Claude Code68
  8. MiMo V2.6 Flash (Reasoning)MiMo Code68
  9. Kimi K2.6 (Reasoning)Kimi Code66
  10. Claude Haiku 4.5 (Reasoning)Claude Code51
MiniMax M3 in MiniMax Code, with entries from other labs in their own coding agents. Results as of 16 November 2026. Source: Earth-Bench V2.

M3 costs just over half of Claude Sonnet 5 per Earth-Bench V2 task, and under a third of Claude Opus 4.7.

It is less careful on GeoBenchX. M3 scores 76 on the solvable tasks, but answers 63% of the impossible ones, against 27% for Claude Sonnet 5. Its GeoBenchX total is 59, against 77. On the Agent Index it scores 68, against 74 for Claude Sonnet 5, at about three quarters of the cost per task.

M3.1 Flash and M3 are not directly comparable here: they run in different harnesses. A Terminus-2 result for M3 will make that comparison possible once it is ranked.

What it costs

Cost per task is what one attempt would cost at the provider's list prices, failed attempts included.

Both MiniMax 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

  • A low-cost open model for spatial questions and databases: M3.1 Flash is level with Nemotron 3 Ultra and Muse Glimmer 30B on GS-QA2 at a fraction of their cost. Put it on your shortlist.
  • Satellite and Earth observation analysis: MiniMax M3 in MiniMax Code is level with Claude Sonnet 5 in Claude Code, at just over half the cost. M3.1 Flash is weaker here.
  • Tasks where a wrong answer is worse than none: M3 answers most impossible GeoBenchX tasks. Add a check before its answers are used.

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.