Google DeepMind Releases Gemini 2.5 Ultra with APAC-Optimised Multilingual Reasoning Benchmarks

Google DeepMind has released Gemini 2.5 Ultra — the company's most capable model to date — with particular emphasis on APAC-language reasoning capabilities that position it competitively against GPT-4o and Claude 3.5 Sonnet for the APAC enterprise AI market. The model achieves state-of-the-art benchmark performance on Japanese reasoning tasks (surpassing GPT-4o on the Japanese JGLUE benchmark by 4.2 points), Mandarin comprehension and generation tasks (achieving 91.3% on the C-Eval benchmark compared to GPT-4o's 87.2%), and Korean language understanding tasks (matching GPT-4o performance on the Korean NLP benchmark suite).

Gemini 2.5 Ultra's APAC-language performance improvements reflect Google's sustained investment in Asian language model training — leveraging Google's substantial corpus of Japanese web content (from Google Search and YouTube), Mandarin content (from non-China Google Search and YouTube), and Korean content — alongside multilingual fine-tuning that specifically targets the reasoning and comprehension tasks most relevant to APAC enterprise use cases: contract analysis in Japanese, financial report summarisation in Mandarin, and customer service response generation in Korean.

The release accompanies expanded Gemini API availability through Google Cloud's APAC regions (Singapore ap-southeast-1, Tokyo asia-northeast1, and Seoul asia-northeast3) — reducing Gemini 2.5 Ultra inference latency for APAC enterprise applications from the 200-400ms that cross-region requests to US regions incur, to the 10-50ms that APAC-region inference provides. For APAC product teams building user-facing AI features where response latency is a UX quality factor, APAC-region Gemini availability makes Gemini 2.5 Ultra practically competitive in latency-sensitive applications for the first time.

For APAC enterprises evaluating AI model providers, Gemini 2.5 Ultra's APAC-language capability improvements create a genuine competitive evaluation between Google, Anthropic, and OpenAI at the frontier — rather than the US-English dominated benchmarks that historically made GPT-4o the default APAC enterprise choice regardless of APAC-language task requirements.