BlogLLM Benchmark Scores as Published, Side by Side
AI2D: Results as Published
Every result the catalog holds for AI2D: 27 models from 8 developers, read from 5 pages. Each number is as printed by the company named beside it; nothing is ranked or rescored. Models are listed by developer, newest first.
Area Multimodal UnderstandingMaintainer Aniruddha Kembhavi and co-authors (Allen Institute for Artificial Intelligence)Introduced 2016Data version 2026103001, as of 2026-10-30
AI2D, AI2 Diagrams, is a set of grade-school science diagrams with multiple-choice questions about what the diagram shows. The paper counts more than 5,000 diagrams and more than 15,000 questions, collected from topics in grade 1 to 6 science textbooks, and each question has four choices. The images are split into 4,000 for training and 1,000 for a blind test. Question answering is scored by accuracy on that test set.
AI2D (test)
| Model | Value | Printed by | Setup as printed |
|---|---|---|---|
| GPT-5.2OpenAI, released 2025-12-11 | 92.2% | Alibaba, Qwen3.5-397B-A17B model cardchecked 2026-10-15 | - |
| GPT-5 miniOpenAI, released 2025-08-07 | 88.2% | Alibaba, Qwen3.5-122B-A10B model cardchecked 2026-10-15 | - |
| GPT-4oOpenAI, released 2024-05-13 | 93.3% | Mistral AI, Medium is the new largechecked 2026-10-15 | - |
| Claude Opus 4.5Anthropic, released 2025-11-24 | 87.7% | Alibaba, Qwen3.5-397B-A17B model cardchecked 2026-10-15 | - |
| Claude Sonnet 4.5Anthropic, released 2025-09-29 | 87% | Alibaba, Qwen3.5-122B-A10B model cardchecked 2026-10-15; printed on 1 more page | - |
| Claude 3.7 SonnetAnthropic, released 2025-02-24 | 78.8% | Mistral AI, Medium is the new largechecked 2026-10-15 | - |
| Gemma 4 26B A4BGoogle, released 2026-04-02 | 88.3% | Alibaba, Qwen3.6-35B-A3B model cardchecked 2026-10-15 | - |
| Gemma 4 31BGoogle, released 2026-04-02 | 89% | Alibaba, Qwen3.6-35B-A3B model cardchecked 2026-10-15 | - |
| Gemini 3 ProGoogle, released 2025-11-18 | 94.1% | Alibaba, Qwen3.5-397B-A17B model cardchecked 2026-10-15 | - |
| Gemma 3 12BGoogle, released 2025-03-12 | 80.4% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit |
| Llama 4 MaverickMeta, released 2025-04-05 | 84.4% | Mistral AI, Medium is the new largechecked 2026-10-15 | - |
| Qwen3.6-35B-A3BAlibaba, released 2026-04-16 | 92.7% | Alibaba, Qwen3.6-35B-A3B model cardchecked 2026-10-15 | - |
| Qwen3.5-122B-A10BAlibaba, released 2026-02-24 | 93.3% | Alibaba, Qwen3.5-122B-A10B model cardchecked 2026-10-15 | - |
| Qwen3.5-27BAlibaba, released 2026-02-24 | 92.9% | Alibaba, Qwen3.5-122B-A10B model cardchecked 2026-10-15; printed on 1 more page | - |
| Qwen3.5-35B-A3BAlibaba, released 2026-02-24 | 92.6% | Alibaba, Qwen3.5-122B-A10B model cardchecked 2026-10-15; printed on 1 more page | - |
| Qwen3.5-397B-A17BAlibaba, released 2026-02-16 | 93.9% | Alibaba, Qwen3.5-397B-A17B model cardchecked 2026-10-15 | - |
| Qwen3-VL-32B-InstructAlibaba, released 2025-10-21 | 85% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit; Context: 32768 |
| 84.8% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit; Context: 4096 | |
| Qwen3-VL-32B-ThinkingAlibaba, released 2025-10-21 | 86.9% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: thinking; Harness: VLMEvalKit; Context: 4096 |
| 87.2% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: thinking; Harness: VLMEvalKit; Context: 40960 | |
| Qwen3-VL-8B-InstructAlibaba, released 2025-10-15 | 82.7% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit; Context: 4096 |
| 83% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit; Context: 32768 | |
| Qwen3-VL-8B-ThinkingAlibaba, released 2025-10-15 | 83.9% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: thinking; Harness: VLMEvalKit; Context: 40960 |
| 83.5% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: thinking; Harness: VLMEvalKit; Context: 4096 | |
| Qwen3-VL-235B-A22BAlibaba, released 2025-09-23 | 89.2% | Alibaba, Qwen3.5-122B-A10B model cardchecked 2026-10-15; printed on 1 more page | - |
| Kimi K2.5Moonshot AI, released 2026-01-27 | 90.8% | Alibaba, Qwen3.5-397B-A17B model cardchecked 2026-10-15 | - |
| Kimi-VL-A3B-InstructMoonshot AI, released 2025-04-10 | 84.6% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit |
| Kimi-VL-A3B-ThinkingMoonshot AI, released 2025-04-10 | 81.2% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: thinking; Harness: VLMEvalKit |
| Mistral Medium 3Mistral AI, released 2025-05-07 | 93.7% | Mistral AI, Medium is the new largechecked 2026-10-15 | - |
| Phi-4-reasoning-vision-15BMicrosoft, released 2026-03-04 | 84.8% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit |
| 84.7% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: force nothink; Harness: VLMEvalKit | |
| 79.7% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: force thinking; Harness: VLMEvalKit | |
| Phi-4-multimodal-instructMicrosoft, released 2025-02-26 | 68.6% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit |
Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.