Google: Gemini 3.1 Flash Lite

Google: Gemini 3.1 Flash Lite is a text-focused model from Google, released on 7 May 2026. It sits in Google's Gemini family and is described by the provider as a generally available, high-efficiency multimodal model built for low-latency, high-volume workloads. If you are already working with other Gemini models, this one slots into the same family but is tuned for efficiency rather than maximum capability.

The model accepts text, image, video, file and audio inputs, and it produces text output only. That means you can send a prompt alongside a screenshot, a PDF, a video clip or an audio recording, and read back a written answer. The context window is 1,048,576 tokens, which is large enough to hold very long documents, extended transcripts or multi-file collections in a single request. The maximum output is 65,536 tokens, so responses can run to substantial length when the task calls for it. Because the model is positioned as a lightweight, low-latency option, it is aimed at workloads where throughput and responsiveness matter as much as raw reasoning depth.

Those capabilities point to a fairly specific set of use cases. High-volume classification, tagging and routing of incoming content is one. Document processing is another: the large context and file input support let you pass contracts, reports or support tickets and ask for summaries or extracted fields. The image, video and audio inputs make it a reasonable fit for media workflows such as captioning, transcription-style extraction or moderation triage, and the provider notes it is designed for lightweight agentic work, so tool calling in simple multi-step pipelines is a natural target. If you need to send the same kind of request thousands of times a day and keep latency low, this model is built for exactly that pattern.

Through api14 you call it with the OpenAI-compatible API using the model id "google/gemini-3.1-flash-lite". It supports include_reasoning, max_tokens, reasoning, reasoning_effort, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools and top_p. In practice that means you can control sampling with temperature and top_p, cap response length with max_tokens, force a shape with response_format or structured_outputs, make runs reproducible with seed, stop generation on chosen strings, and attach tools with tools and tool_choice when you want the model to call functions. The reasoning and reasoning_effort parameters, along with include_reasoning, let you tune how much internal reasoning is exposed or requested. If your task depends on very deep multi-step reasoning, a larger model in the same family may fit better, but for efficient multimodal processing at scale this is the intended choice.