The Shiro Corporation - The giant technology company Google officially introduced the new member of its artificial intelligence model ( artificial intelligence The newest addition to the Gemini 2.5 series family is the Gemini 2.5 Flash Lite, on Tuesday (17/6/2025).
Google calls Gemini 2.5 Flash Lite the fastest AI model among the Gemini series with the lowest power consumption.
"We have also introduced a preview for the Gemini 2.5 Flash Lite model, the most fast and cost-efficient 2.5 model. You can start developing with the preview of the latest model and we look forward to your feedback and suggestions," said Senior Director of Product Management Gemini Tulsee Doshi in an official Google blog.
This model is not yet available to the public and is still in development (beta version), but Gemini 2.5 Flash Lite is already accessible to developers. developers ) through Google AI Studio or Vertex AI platform.
Referring to Google's announcement on its official blog, Gemini 2.5 Flash Lite is able to surpass the scores of its predecessor, Gemini 2.0 Flash Lite, in some tests. benchmarking .
Starting from programming tests, completing math tasks, science, reasoning logic, to multimodal assignments.
Google mentions that for specific testing of GPQA (Graduate-Level Google-Proof Q7A Benchmark/problem-solving benchmark for science), AIME (American Invitational Mathematics Examination), as well as LiveCodeBench (measurement of LLM capabilities in coding contexts), scores are high.
The test scores achieved consistently surpassed the Gemini 2.0 Flash Lite scores. Even for some tests, such as visual reasoning, multi-lingual performance, and AIME 2025, the scores were nearly in tight competition with the Gemini 2.5 Flash models.
Another advantage possessed by Gemini 2.0 Flash Lite is its low cost consumption. In Standard and Thinking modes, the cost required is approximately $0.10 (IDR 1,631) per million tokens and $0.40 (IDR 6,527) per million output tokens.
However, as quoted KompasTekno dari The Decoder , Wednesday (18/6/2025), the cost details of tokens issued for Thinking mode will be higher. To generate more accurate presentations, the actual usage costs will also follow.
Therefore, Gemini 2.5 Flash Lite is suitable for handling high-weight and low-latency tasks, such as translating text from one language to another, up to classification.
Because, testing scores for the two tasks above shows the results. benchmark high, namely 86.8 percent for the FACTS Grounding questions and 84.5 percent for the Multilingual MMLU.
FACTS Grounding in AI testing refers to the ability of large language models (LLMs) to provide responses based on current information, with sources that can be verified, thereby reducing the hallucination of information.
Meanwhile, Multilingual MMLU refers to the testing metrics used to evaluate general knowledge and reasoning ability of language models across different languages and domains.
The other testing, such as visual reasoning (Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark/MMMU) for Gemini 2.5 Flash Lite recorded a high score of 72.9 percent, while visual reasoning and image understanding scored 57.5 percent.
Like other models in the Gemini 2.5 series, this supposedly most efficient model also supports integrated multi-modal inputs, such as Google Search, code execution, and understanding of context.
Google 2.5 Flash and Gemini 2.5 Pro
In conjunction with this launch, Google also announced that the Gemini 2.5 Flash and Pro AI models are now officially accessible to the public. Both models have successfully achieved a score benchmark which is considered high for some tests.
This official release was finally carried out after Google made a number of developments and improvements. Both models are also available on both Android and iOS through the Gemini app.
All members of the Gemini 2.5 family, which were developed, are designed for hybrid reasoning, i.e., systems that combine various approaches to problem-solving. This ranges from logical reasoning to statistical methods.
This integration is capable of helping AI models generate accountable and optimal solutions for complex tasks.
This integration strategy is also aimed at balancing high performance with low cost and latency. Google is also positioning the three Gemini 2.5 models as part of the pareto front.
“We designed Gemini 2.5 as a family of hybrid reasoning models with outstanding performance, while also applying the Pareto Front to prioritize speed and cost efficiency,” Tulsee explained.
The Pareto Front in the context of AI is the ability to provide optimal solutions for multi-objective optimization problems. Simply put, the way this model works is by improving one aspect at the expense of another.
This model allows users to obtain the most effective solution or one that best meets their needs or priorities among conflicting objectives. The Pareto principle is applied to optimize the efficiency and capability of the released model.
Comments
Post a Comment