RefCOCO average (RefCOCO (avg))
We show this table for reference; we do not rank on it.
A referring-expression grounding benchmark averaged across RefCOCO variants to test whether a model can localize described objects correctly.
Benchmark score on RefCOCO (avg) — November 13, 2026
We compile the RefCOCO (avg) rows from provider self-reports and secondary reports. Qwen3.6-27B leads the table at 92.5%, followed by Qwen3.5 397B (92.3%) and Qwen3.6-35B-A3B (92.0%). We do not use these results to rank models overall.
Qwen3.6-27B
Alibaba
Reported by AlibabaQwen3.5 397B
Alibaba
Reported by AlibabaQwen3.6-35B-A3B
Alibaba
Reported by Alibaba10 modelsMultimodal & GroundedCurrentDisplay onlyUpdated November 13, 2026
Benchmark score table (10 models)
None of the 10 results on this page has a documented run by the benchmark or an independent evaluator: 8 reported by the labs that built the models and 2 reported in another publisher’s document.
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Qwen3.6-27BAlibabaReported by Alibaba | 92.5% | Reported by Alibaba | Not reported | Open |
| 2 | Qwen3.5 397BAlibabaReported by Alibaba | 92.3% | Reported by Alibaba | Not reported | Open |
| 3 | Qwen3.6-35B-A3BAlibabaReported by Alibaba | 92.0% | Reported by Alibaba | Not reported | Open |
| 4 | Nemotron 3 Nano Omni 30B A3BNVIDIAReported by NVIDIA | 90.5% | Reported by NVIDIA | Not reported | Open |
| 5 | LFM3-VL-2BLiquidAIReported by LiquidAI | 90.5% | Reported by LiquidAI | Not reported | Open |
| 6 | LFM2.5-VL-3BLiquidAIReported by LiquidAI | 87.9% | Reported by LiquidAI | Not reported | Open |
| 7 | Kimi K2.5Moonshot AIAnother lab’s comparison table | 87.8% | Another lab’s comparison table | Not reported | Open |
| 8 | ZAYA1-VL-8BZyphraReported by Zyphra | 84.3% | Reported by Zyphra | Not reported | Open |
| 9 | Gemini 3 ProGoogleAnother lab’s comparison table | 84.1% | Another lab’s comparison table | Not reported | Closed |
| 10 | Interfaze BetaInterfazeReported by Interfaze | 82.1% | Reported by Interfaze | Not reported | Closed |
- Lab-reported
- Benchmark-run
- Independent
Among the reported RefCOCO (avg) rows, Qwen3.6-27B is first at 92.5%. The third row is 0.5 points behind. The broader top-10 range is 10.4 points, so the table still separates the published systems.
10 models have been evaluated on RefCOCO (avg). The benchmark falls in the Multimodal & Grounded category. RefCOCO (avg) is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About RefCOCO (avg)
Year
2026
Tasks
Referring-expression grounding
Format
Grounded visual localization
Difficulty
Fine-grained visual grounding
RefCOCO-style tasks matter for grounding-heavy assistants because they measure whether the model can map language to specific objects or regions instead of only answering abstract questions. BenchLM stores provider-reported aggregate RefCOCO values as a display-only grounding row.
Freshness and provenance
Version
RefCOCO (avg) 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 RefCOCO (avg) measure?
A referring-expression grounding benchmark averaged across RefCOCO variants to test whether a model can localize described objects correctly.
Which model scores highest on RefCOCO (avg)?
Qwen3.6-27B by Alibaba currently leads with a score of 92.5% on RefCOCO (avg).
How many models are evaluated on RefCOCO (avg)?
10 AI models have published results on RefCOCO (avg) in the BenchLM catalog.
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.