GroundingME: Multi-Dimensional Visual Grounding (GroundingME)
We show this table for reference; we do not rank on it.
A visual grounding benchmark of 1,005 examples that tests whether multimodal models localize objects across four dimensions: discriminative look-alikes, spatial relations, occluded or tiny targets, and ungroundable rejection queries.
Accuracy@0.5 on GroundingME — GroundingME CVPR 2026 leaderboard (captured November 20, 2026)
We mirror the published accuracy@0.5 view for GroundingME. Qwen3-VL-A22B (Thinking) leads the public snapshot at 49.8%, followed by Qwen3-VL-32B (Thinking) (46.9%) and Seed-1.6-Vision-250815 (Thinking) (46.5%). We do not use these results to rank models overall.
Qwen3-VL-A22B (Thinking)
Alibaba
Open-source · Dis 65.2 · Spa 73.7 · Lim 45.0 · Rej 5.5
Qwen3-VL-32B (Thinking)
Alibaba
Open-source · Dis 65.7 · Spa 70.0 · Lim 36.0 · Rej 9.5
Seed-1.6-Vision-250815 (Thinking)
ByteDance
Commercial · Dis 59.3 · Spa 72.7 · Lim 41.7 · Rej 1.5
32 modelsMultimodal & GroundedCurrentDisplay onlyUpdated GroundingME CVPR 2026 leaderboard (captured November 20, 2026)
Accuracy@0.5 table (32 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Qwen3-VL-A22B (Thinking)AlibabaOpen-source · Dis 65.2 · Spa 73.7 · Lim 45.0 · Rej 5.5 | 49.8% | Not reported | Unknown |
| 2 | Qwen3-VL-32B (Thinking)AlibabaOpen-source · Dis 65.7 · Spa 70.0 · Lim 36.0 · Rej 9.5 | 46.9% | Not reported | Unknown |
| 3 | Seed-1.6-Vision-250815 (Thinking)ByteDanceCommercial · Dis 59.3 · Spa 72.7 · Lim 41.7 · Rej 1.5 | 46.5% | Not reported | Unknown |
| 4 | Qwen3-VL-235B-A22B InstructAlibabaOpen-source · Dis 69.6 · Spa 49.7 · Lim 54.0 · Rej 0.0 | 45.1% | Not reported | Open |
| 5 | Seed-1.6-Vision-250815ByteDanceCommercial · Dis 59.8 · Spa 58.7 · Lim 42.7 · Rej 1.0 | 42.6% | Not reported | Unknown |
| 6 | Qwen3-VL-32BAlibabaOpen-source · Dis 75.0 · Spa 47.3 · Lim 34.0 · Rej 0.0 | 39.5% | Not reported | Unknown |
| 7 | Qwen3-VL-A3B (Thinking)AlibabaOpen-source · Dis 53.4 · Spa 53.3 · Lim 38.0 · Rej 5.5 | 39.2% | Not reported | Unknown |
| 8 | Qwen3-VL-A3BAlibabaOpen-source · Dis 63.2 · Spa 30.0 · Lim 46.7 · Rej 0.0 | 35.7% | Not reported | Unknown |
| 9 | Qwen3-VL-8B (Thinking)AlibabaOpen-source · Dis 52.5 · Spa 43.0 · Lim 33.3 · Rej 4.5 | 34.3% | Not reported | Unknown |
| 10 | GLM-4.5V (Thinking)Z.aiOpen-source · Dis 52.5 · Spa 45.3 · Lim 30.3 · Rej 4.0 | 34.0% | Not reported | Unknown |
| 11 | Qwen3-VL-4BAlibabaOpen-source · Dis 56.4 · Spa 28.3 · Lim 47.0 · Rej 0.0 | 33.9% | Not reported | Unknown |
| 12 | GLM-4.5VZ.aiOpen-source · Dis 52.9 · Spa 42.0 · Lim 29.3 · Rej 0.5 | 32.1% | Not reported | Unknown |
| 13 | Qwen3-VL-8BAlibabaOpen-source · Dis 61.3 · Spa 26.3 · Lim 36.0 · Rej 0.0 | 31.0% | Not reported | Unknown |
| 14 | Qwen2.5-VL-72BAlibabaOpen-source · Dis 48.5 · Spa 40.3 · Lim 23.7 · Rej 3.0 | 29.6% | Not reported | Unknown |
| 15 | Qwen2.5-VL-32BAlibabaOpen-source · Dis 47.5 · Spa 40.0 · Lim 17.7 · Rej 0.0 | 26.9% | Not reported | Open |
| 16 | MiMo-VL-7B-RL-2508 (Thinking)XiaomiOpen-source · Dis 46.6 · Spa 28.7 · Lim 17.0 · Rej 5.0 | 24.1% | Not reported | Unknown |
| 17 | Qwen3-VL-2BAlibabaOpen-source · Dis 44.6 · Spa 11.7 · Lim 28.7 · Rej 0.0 | 21.1% | Not reported | Unknown |
| 18 | Gemini 2.5 ProGoogleCommercial · Dis 34.8 · Spa 34.0 · Lim 7.0 · Rej 7.0 | 20.7% | Not reported | Closed |
| 19 | Gemini 2.5 FlashGoogleCommercial · Dis 36.3 · Spa 25.0 · Lim 13.0 · Rej 0.0 | 18.7% | Not reported | Closed |
| 20 | MiMo-VL-7B-RL-2508XiaomiOpen-source · Dis 44.1 · Spa 19.3 · Lim 13.0 · Rej 0.0 | 18.6% | Not reported | Unknown |
| 21 | InternVL3.5-A28BOpenGVLabOpen-source · Dis 28.4 · Spa 25.0 · Lim 13.0 · Rej 0.0 | 17.1% | Not reported | Unknown |
| 22 | Qwen2.5-VL-7BAlibabaOpen-source · Dis 31.9 · Spa 14.3 · Lim 14.3 · Rej 0.5 | 15.1% | Not reported | Unknown |
| 23 | Llama 4 MaverickMetaOpen-source · Dis 18.1 · Spa 22.3 · Lim 5.0 · Rej 6.0 | 13.0% | Not reported | Open |
| 24 | Llama-Nemotron-8BNVIDIAOpen-source · Dis 25.0 · Spa 6.0 · Lim 8.3 · Rej 5.5 | 10.4% | Not reported | Unknown |
| 25 | Llama 4 ScoutMetaOpen-source · Dis 17.6 · Spa 12.3 · Lim 3.7 · Rej 2.5 | 8.9% | Not reported | Open |
| 26 | Keye-VL-1.5-8BKwaiOpen-source · Dis 21.6 · Spa 8.0 · Lim 5.7 · Rej 0.0 | 8.5% | Not reported | Unknown |
| 27 | LLaVA-OneVision-1.5-8BLLaVAOpen-source · Dis 9.8 · Spa 4.7 · Lim 3.3 · Rej 0.0 | 4.4% | Not reported | Unknown |
| 28 | MiniCPM-V-4.5OpenBMBOpen-source · Dis 7.8 · Spa 4.0 · Lim 4.0 · Rej 0.0 | 4.0% | Not reported | Unknown |
| 29 | InternVL3.5-8BOpenGVLabOpen-source · Dis 6.4 · Spa 4.0 · Lim 1.7 · Rej 1.5 | 3.3% | Not reported | Unknown |
| 30 | Mistral-3.2-24BMistralOpen-source · Dis 4.4 · Spa 2.7 · Lim 0.0 · Rej 0.0 | 1.7% | Not reported | Unknown |
| 31 | Gemma 3 27BGoogleOpen-source · Dis 1.5 · Spa 0.3 · Lim 0.0 · Rej 0.0 | 0.4% | Not reported | Open |
| 32 | Phi-4 Multimodal InstructMicrosoftOpen-source · Dis 1.0 · Spa 0.7 · Lim 0.0 · Rej 0.0 | 0.4% | Not reported | Open |
Results reported by model developersUnranked
Labs ran these themselves. They take no rank here.
| Model | Reported value | What was measured | Against the official board | Who ran it |
|---|---|---|---|---|
| Inkling 2 MiniThinking Machines Lab | 55.5%Overall accuracy | Reported in the Inkling 2 Mini model cardNot yet evaluated on the official GroundingME CVPR 2026 leaderboard. | Comparability not established | Reported by Thinking Machines Lab |
A lab’s own result is shown for reference. It never takes a rank on this board.
How we label sourcesHow to read this leaderboard
Operator receipt: 32 sourced rows are currently displayable on this page; the leading published row is Qwen3-VL-A22B (Thinking) at 49.8%.
Honest limit: Benchmark dataset items follow SA-1B and HR-Bench research-only terms; we mirror only published aggregate scores. Newer lab-reported rows are kept separate because bounding-box prompting and IoU handling can vary across labs.
How we show GroundingME
We mirror the GroundingME CVPR 2026 leaderboard from Peking University and Xiaomi: 32 open-source and commercial vision-language models evaluated across 1,005 grounding examples.
Each row shows overall Accuracy@0.5 (IoU >= 0.5) alongside the four capability dimensions reported on the project page: Discriminative look-alikes, Spatial relations, Limited visibility, and Rejection of ungroundable queries.
We mirror only published aggregate model scores and exclude GroundingME from overall and category rankings. Provider-reported results not yet on the official board appear below under lab-reported results.
Snapshot
The published GroundingME snapshot places Qwen3-VL-A22B (Thinking) first at 49.8%. The third row is 3.3 points behind. The broader top-10 range is 15.8 points, so the table still separates the published systems.
32 models have been evaluated on GroundingME. The benchmark falls in the Multimodal & Grounded category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. GroundingME is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About GroundingME
Year
2025
Tasks
1,005 multi-dimensional visual grounding examples
Format
Accuracy@0.5 (%)
Difficulty
Fine-grained object localization and ungroundable rejection
Introduced by Li et al. (CVPR 2026) from Peking University and Xiaomi. We mirror the official CVPR 2026 leaderboard (32 models: 28 open-source and 4 commercial) with overall Accuracy@0.5 and the four sub-dimension scores (Discriminative, Spatial, Limited, Rejection). Provider-reported results for newer models not yet on the official board appear below under lab-reported results.
Freshness and provenance
Version
GroundingME 2025
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 GroundingME measure?
A visual grounding benchmark of 1,005 examples that tests whether multimodal models localize objects across four dimensions: discriminative look-alikes, spatial relations, occluded or tiny targets, and ungroundable rejection queries.
Which model leads the published GroundingME snapshot?
Qwen3-VL-A22B (Thinking) currently leads the published GroundingME snapshot with 49.8% accuracy@0.5. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on GroundingME?
The GroundingME CVPR 2026 leaderboard (captured November 20, 2026) snapshot contains 32 AI models.
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.