ByteDance: UI-TARS 7B

ByteDance: UI-TARS 7B demonstrates strong overall performance, particularly in operational efficiency. It performs among the fastest models, ranking in the top tier for speed (75th percentile), and consistently offers highly competitive pricing (81st percentile). The model also exhibits exceptional reliability with a 99% success rate across benchmarks, indicating minimal technical failures. In terms of specific benchmark performance, UI-TARS 7B shows notable strengths in ethical reasoning and general knowledge, achieving 99.0% and 93.0% accuracy respectively. It also performs well in mathematics (78.0%) and coding (68.0%). However, the model exhibits significant weaknesses in hallucination mitigation, with only 64.0% accuracy, suggesting a tendency to generate responses for fictional concepts. Its performance in email classification (82.0%), instruction following (32.3%), and complex reasoning (36.0%) is also below average compared to other models. Despite these areas for improvement, its core strength lies in its multimodal vision-language agent capabilities for GUI-based environments, as highlighted by its state-of-the-art results on interactive and grounding benchmarks like OSworld and WebVoyager, and perfect task completion in Poki games.