google / gemini-3.5-flash-lite

## Overview Gemini 3.5 Flash-Lite is a natively multimodal reasoning model made by Google. It takes text plus optional images, video, audio, and PDFs, then returns text output. It’s tuned for high-throughput tasks like document parsing, translation, and classification, where you need consistent, short answers at scale. It also supports structured JSON outputs, which helps you extract fields into a predictable shape. ## What you can build - Receipt and invoice extraction from PDFs into JSON - High-volume classification for support tickets, chats, or emails - Document parsing for contracts, policies, and technical PDFs up to 1,000 pages - Audio transcription, highlights, and timeline summaries for long recordings - Video summaries with timestamp references and key moment extraction - Routing “sub-agent” steps that decide what tool or workflow should run next ## Inputs - A required instruction written as plain text. Put the task, constraints, and output format rules here. - Optional file attachments, up to 50 total per run, when you need multimodal understanding. - Optional images in PNG, JPEG, WEBP, HEIC, or HEIF format. - Optional videos in common formats like MP4, MOV, AVI, WebM, or WMV. - Optional audio in WAV, MP3, AIFF, AAC, OGG Vorbis, or FLAC format. - Optional PDFs (application/pdf) for document understanding. Gemini can handle documents up to 1,000 pages. - Optional system-level instructions that set global behavior for the response. - Optional conversation identifiers (a user ID and a session ID) if you want chat history continuity. - Optional reasoning effort setting. Higher effort helps multi-step tasks and tool-heavy flows. - Optional sampling controls to change response randomness. These are advanced and often unnecessary. - An optional maximum response length cap, up to 65,536 output tokens. ## Outputs The model returns text. - For chat and summarization, you get a natural-language answer. - For extraction, you can request structured output that conforms to a provided JSON Schema subset. - For multimodal inputs, the text can include grounded references like timestamps for video. ## Recommended settings - Use minimal reasoning effort for extraction, routing, and classification at high volume. - Use medium or high reasoning effort for multi-step sub-agent tasks, code execution, or tool calling. - For extraction work, request structured JSON output so downstream parsing stays reliable. ## Limitations - Output is text only. Image generation and audio generation are not supported. - Computer-use style automation is not supported. - The input context limit is 1,048,576 tokens. The output limit is 65,536 tokens. - The knowledge cutoff date is March 2026. It may miss newer events. - Like other foundation models, it can hallucinate facts. Verify critical details. - Video understanding samples frames. Fast action can lose detail. - Audio is downsampled and multi-channel audio is merged. Speaker separation can degrade. - PDFs work best when pages are clean and readable. Scanned, blurry, or skewed pages raise error rates. - Non-PDF documents may be treated as plain text. This can drop charts and layout context. ## Safety & compliance Gemini API includes adjustable safety filters for four categories. - Harassment - Hate speech - Sexually explicit content - Dangerous content Extra safety filters are off by default for Gemini 3 models, but you can enable stricter blocking when needed. Some harms are always blocked and can’t be relaxed. This includes child safety related content. Google reports that Gemini 3.5 Flash-Lite met required launch thresholds for child safety in human red teaming. Test your prompts on real user traffic patterns. Tighten filters if you see borderline content.