AI2D test split (AI2D_TEST)
We show this table for reference; we do not rank on it.
A diagram understanding benchmark focused on scientific and educational visual question answering.
Benchmark score on AI2D_TEST — November 13, 2026
We compile the AI2D_TEST rows from provider self-reports and secondary reports. Gemini 3 Pro leads the table at 94.1%, followed by Qwen3.5 397B (93.9%) and Qwen3.6-35B-A3B (92.7%). We do not use these results to rank models overall.
Gemini 3 Pro
Qwen3.5 397B
Alibaba
Reported by AlibabaQwen3.6-35B-A3B
Alibaba
Reported by Alibaba8 modelsMultimodal & GroundedCurrentDisplay onlyUpdated November 13, 2026
Benchmark score table (8 models)
None of the 8 results on this page has a documented run by the benchmark or an independent evaluator: 4 reported by the labs that built the models and 4 reported in another publisher’s document.
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Gemini 3 ProGoogleAnother lab’s comparison table | 94.1% | Another lab’s comparison table | Not reported | Closed |
| 2 | Qwen3.5 397BAlibabaReported by Alibaba | 93.9% | Reported by Alibaba | Not reported | Open |
| 3 | Qwen3.6-35B-A3BAlibabaReported by Alibaba | 92.7% | Reported by Alibaba | Not reported | Open |
| 4 | GPT-5.2OpenAIAnother lab’s comparison table | 92.2% | Another lab’s comparison table | Not reported | Closed |
| 5 | Kimi K2.5Moonshot AIAnother lab’s comparison table | 90.8% | Another lab’s comparison table | Not reported | Open |
| 6 | Nemotron 3 Nano Omni 30B A3BNVIDIAReported by NVIDIA | 88.5% | Reported by NVIDIA | Not reported | Open |
| 7 | Claude Opus 4.5AnthropicAnother lab’s comparison table | 87.7% | Another lab’s comparison table | Not reported | Closed |
| 8 | ZAYA1-VL-8BZyphraReported by Zyphra | 87.5% | Reported by Zyphra | Not reported | Open |
- Lab-reported
- Benchmark-run
- Independent
Among the reported AI2D_TEST rows, Gemini 3 Pro is first at 94.1%. The third row is 1.4 points behind. The broader top-10 range is 6.6 points, so many of the published results sit in a relatively narrow band.
8 models have been evaluated on AI2D_TEST. The benchmark falls in the Multimodal & Grounded category. AI2D_TEST is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About AI2D_TEST
Year
2026
Tasks
Diagram understanding
Format
Diagram-grounded QA
Difficulty
Structured visual reasoning
AI2D-style tasks matter because diagrams compress structure differently from photos or office documents. They test whether a model can parse arrows, labels, and spatial relations in technical illustrations.
Freshness and provenance
Version
AI2D_TEST 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does AI2D_TEST measure?
A diagram understanding benchmark focused on scientific and educational visual question answering.
Which model scores highest on AI2D_TEST?
Gemini 3 Pro by Google currently leads with a score of 94.1% on AI2D_TEST.
How many models are evaluated on AI2D_TEST?
8 AI models have published results on AI2D_TEST in the BenchLM catalog.
Compare top models on AI2D_TEST
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.