AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。
Google LLC today launched its most capable entry-level artificial intelligence model yet. Gemini 3.7 Flash is rolling out three weeks after its predecessor. Despite the short release cycle, Google engineers managed to implement significant output quality improvements. The company says that Gemini 3.7 Flash outperformed comparable models from Anthropic PBC and OpenAI Group PBC across nine benchmarks. One of the evaluations in which the AI earned first place is FrontierCode 1.1 Main. It comprises 100 programming tasks spanning multiple languages. The benchmark requires models to not only produce working code, but also comply with other requirements that often crop up in enterprise software projects. Code must undergo bug testing before it’s submitted and follow project-specific style guides. Compared with its predecessor, Gemini 3.7 is particularly adept at generating user interfaces. Google says layouts designed by the model more closely align with reference images uploaded by the user. Gemini 3.7 Flash can process up to 1 million tokens worth of images, video and text per prompt. Its prompt responses comprise up to 64,000 tokens of text. “Gemini 3.7 Flash delivers a noticeably improved developer experience over 3.6 Flash,” Tulsee Doshi, senior director of product management at Google, wrote in a blog post. “It better adapts to roadblocks, clarifies intent when needed, and follows instructions with greater fidelity. It thinks more diligently, putting in more effort into multi-step planning and tool calls.” The model also lends itself to other tasks besides programming. Google tested it using GDP.pdf benchmark, a benchmark that requires neural networks to answer questions about business documents. It answered 34% of the questions correctly, which put it 6% and 9.3% ahead of Claude Sonnet 5 and GPT-5.6 Terra, respectively. Furthermore, Google says that Gemini 3.7 Flash can power AI agent ensembles. Google didn’t specify the model’s architecture. Gemini 3.7 Flash’s model card indicates that it’s based on the same architecture as the company’s previous-generation Gemini 3.6 Flash model. That algorithm, in turn, is derived from Gemini 3 Pro, which features a transformer-based mixture-of-experts architecture. Gemini 3.7 Flash’s model card also lacks information about how it was trained. Developers often train entry-level neural networks by distilling the output of a more capable AI in the same algorithm series. Meta Platforms Inc., for example, created its recently released Muse Glimmer language model by distilling Muse Spark. Google will enable developers to access Gemini 3.7 Flash for half the price of Gemini 3.6 Flash through the end of the year. Additionally, the company is bringing the model to Gemini Spark, a consumer-focused AI agent that debuted in March. It can browse the web and perform actions in other Google services. Image: Google