Google: Gemini 3 Flash Preview (batch)

Gemini 3 Flash Preview is a Google reasoning model for multimodal understanding, coding and agentic tasks. It accepts text, images, video, audio and PDF and returns text. A request can therefore join written instructions, source files, visual observations and conversation history around one problem. The result may be an explanation, a code proposal, extracted data or a plan grounded in those materials, rather than a generated image or spoken response.

The exact model card gives an input limit of 1,048,576 tokens and an output limit of 65,536 tokens. Plan the source and result separately: a long repository excerpt, document collection or recording can fit only when its processed representation and instructions fit the input budget, and the answer still needs its own allotted length. Label related sources and identify which material is authoritative. For a long prompt, the developer guide recommends placing the specific question after the data and tying it explicitly to the preceding information.

Thinking is dynamic, with high as the documented default for this model. The supported levels are minimal, low, medium and high. They are relative reasoning allowances rather than guaranteed token counts. Minimal can still involve some reasoning on a difficult coding problem. Choose a level according to the task: a short transformation and a multi-condition debugging problem need different treatment. Use direct, clear instructions instead of forcing elaborate reasoning rituals into the prompt; request the final artifact and the checks that make it usable.

The producer card lists structured outputs, function calling, code execution, computer use, file search, search grounding, Google Maps grounding, URL context and caching. Built-in tools can be combined with custom function calling in the documented Gemini 3 flow. An application defines the functions, checks their arguments, executes accepted actions and sends back their results. For a code assistant, a read-file or controlled test-runner tool can provide concrete evidence for the next step. A proposed action and an observed tool result should remain distinct in the final report.

For programming, supply the relevant files, language, expected behavior and an input that demonstrates the issue. Set the scope of the change and name the interface or behavior that must remain. Ask for a revised implementation and the cases that should be checked. For an analytical document task, define the question, required fields and source references. If structured output is used, provide a schema with clear field meanings and missing-value rules. Validate both its syntax and whether each value follows from the original material.

For multimodal analysis, identify what matters in each source. A screenshot can support an interface explanation, a PDF can supply a table or chart, and a recording can provide spoken or visible events. The developer guide discusses media-resolution controls because detail changes the processed representation and token use. The producer’s code-execution workflow can crop or zoom supplied images, run calculations from visual data and add annotations before the model prepares its answer. Readable small text may require a different setting from a broad scene description. Keep the relevant page, image or clip available so observations can be checked rather than treating a fluent summary as evidence that every detail was read correctly.

Conversation and tool state need to be retained in the chosen interface. Gemini 3 uses thought signatures to carry reasoning context between calls; the developer guide distinguishes managed stateful history from manually managed stateless history. Preserve the required context when continuing a tool task or revising a result. Google also recommends keeping its default temperature of 1.0 rather than importing a low-temperature recipe that can cause loops or degraded behavior on complex tasks. These are producer integration practices, not new claims about any particular application’s transport.

The text model does not advertise image or audio generation or the Live API. The guide also distinguishes visual understanding from pixel-level image segmentation, which is not a supported Gemini 3 Flash capability. Use an appropriate final result for the actual task: source-linked observations, a structured record, a code change or a decision with its assumptions stated. Check arithmetic, verify code in the intended environment and review returned tool information. When recent facts matter, use the configured grounding sources or supplied current data instead of assuming that model knowledge is a live database.