google / gemini-3.8-flash

Overview

Gemini 3.8 Flash is a Google DeepMind model built for long-horizon reasoning and agent workflows. (deepmind.google)

You give it text prompts and, if needed, files like images, video, audio, or PDFs. It reads those inputs and produces a text response. (ai.google.dev)

It supports very long context. The Gemini API lists a 1,048,576 token input limit and a 65,536 token output limit for this model. (ai.google.dev)

Gemini 3.8 Flash can work with built-in capabilities like code execution, function calling, file search, URL context, Search grounding, and Google Maps grounding. It does not support image generation or audio generation. (ai.google.dev)

What you can build

• Long-horizon coding helpers that plan changes, reason over diffs, and produce patch-ready steps
• Agentic knowledge workflows that combine your files with grounded web results
• PDF analysis for contracts, reports, and slide decks, including table and chart Q&A
• Video understanding for tutorials, product demos, and incident recordings with detailed summaries
• Audio understanding for meeting summaries, call notes, and transcript-based Q&A
• Multimodal support bots that answer questions from screenshots plus a PDF manual
• Structured data extraction flows that return JSON-shaped results when a strict format is required (ai.google.dev)

Inputs

• A required instruction or question you want answered as plain text.
• Optional attachments (up to 50 per request in this Wiro run). Attach what the model should read, such as: - Images (BMP, JPEG, PNG, WebP). (ai.google.dev)
  • Videos (MP4, MPEG, MOV/QuickTime, AVI, WebM, WMV, FLV, 3GPP). (ai.google.dev)
  • Audio (WAV, MP3, AIFF, AAC, OGG, FLAC, M4A, Opus, WebM, and others listed by the Gemini API docs). (ai.google.dev)
  • PDFs for document understanding. (ai.google.dev)
• If your workflow sends files inline, keep payload size within Gemini’s inline limits (100 MB per request, or 50 MB for PDFs). (ai.google.dev)
• Optional system-level instructions to set rules and tone for the whole session.
• Optional thinking level selection (low, medium, or high). Gemini 3.8 Flash rejects the “minimal” level. (ai.google.dev)
• Optional grounding toggles: - Ground the response with Google Search results.
  • Ground the response with Google Maps place data. These are model capabilities and must be enabled by the workflow. (ai.google.dev)
• Optional response length control up to the model’s max output limit.
• Optional user and session identifiers when you want chat history to persist across turns.

Outputs

• A text response. It can include paragraphs, lists, tables, and code blocks.
• When the workflow requests structured output, the model can produce a JSON response that matches a provided JSON Schema. This is designed for type-safe extraction and downstream automation. (ai.google.dev)

Limitations

• Output is text-only. The model does not generate images or audio. (ai.google.dev)
• The model can hallucinate. Treat unverified facts as suggestions, not guarantees. (deepmind.google)
• Knowledge can be stale for some topics. The model card lists a knowledge cutoff of March 2026, with some domains limited to January 2025. (deepmind.google)
• Video understanding may miss fast action details in default processing. Gemini’s static mode extracts frames at 1 FPS and processes audio at low bitrate, which can lose fine-grained motion cues. (ai.google.dev)
• PDFs have hard limits. Gemini’s document guide describes support up to 50 MB or 1,000 pages, and it scales pages to fit max resolutions during processing. (ai.google.dev)
• Low-quality inputs increase errors. Blurry scans, rotated pages, noisy audio, and heavily compressed video reduce accuracy. (ai.google.dev)

Safety & compliance

Google applies safety mitigations and content policies to Gemini 3.8 Flash. The model may refuse requests for disallowed content, including harmful instructions. (deepmind.google)

Don’t upload sensitive personal data unless you have rights and consent. Review outputs before you act on them, especially for security, legal, or medical decisions. (deepmind.google)