Moondream3.1 Release Comparators - RefCOCO-M (val): leaderboard
Moondream July 7, 2026 release comparator evaluation. Source prints Qwen3.5 9B, LocateAnything and Gemma4 12B; LocateAnything explicitly uses slow mode to maximize accuracy; vLLM, one H100 SXM5, batch size16, highest accuracy and reasoning settings chosen for each model. Exact checkpoint revisions, precision and reasoning token budgets are not disclosed; keep source-qualified identities and do not infer Max/High/BF16 or merge Base/Instruct variants. Detection metric explicitly F1@0.5. Moondream values are withheld because per-task Lens LoRA versus distilled-base assignment is unresolved; this does not affect these separately printed comparator measurements. No throughput or mixed-average capability board. Metric: F1@0.5 (%). Higher is better. Source: https://moondream.ai/blog/moondream-3-1-beyond-benchmarks
Metric: F1@0.5 (%). Source: moondream.ai. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3.5 9B [Moondream3.1 release baseline] | 85.73 | |
| 2 | Gemma 4 12B [Moondream3.1 release baseline] | 74.04 | |
| 3 | LocateAnything [Moondream3.1 release baseline, slow mode] | 46.21 |
Interactive version: theaggregate.ai/benchmark?slug=moondream3-1-release-comparators-refcoco-m-val · How It Works · Data refreshed daily, snapshot 2026-10-17.