Google: Gemini 3 Flash Preview

Gemini 3 Flash is optimized for extended agentic workflows where the model maintains context across multiple turns while dynamically calling external tools. It uses a stateless request-response pattern where each turn includes full conversation history, tool definitions via JSON schema, and execution results, enabling the model to reason about tool outputs and decide next actions without server-side session management.

Gemini 3 Flash generates code across 40+ programming languages using a transformer-based approach that understands syntax, semantics, and common patterns. It supports streaming output (token-by-token delivery) for real-time IDE integration, and accepts multi-file context to generate code aware of existing codebase structure, imports, and dependencies without requiring explicit AST parsing.

Gemini 3 Flash accepts and processes multiple input modalities in a single request: text prompts, images (JPEG, PNG, WebP, GIF), audio files (MP3, WAV, etc.), and video frames. The model uses a unified embedding space where all modalities are converted to token representations, allowing it to reason across modalities (e.g., describe an image, transcribe audio, or answer questions about video content) without separate preprocessing pipelines.

Gemini 3 Flash can extract structured data from unstructured text or images by accepting a JSON Schema definition of the desired output format. The model constrains its output to match the schema, returning valid JSON that can be directly parsed without post-processing. This works via a constrained decoding approach where the model's token generation is guided by the schema to ensure type correctness and required field presence.

Gemini 3 Flash supports server-sent events (SSE) streaming where tokens are delivered one-by-one as they are generated, enabling real-time display in client applications. The streaming protocol includes metadata for each token (finish reason, safety ratings) and supports cancellation mid-stream. This allows applications to display model output character-by-character without waiting for full response completion, reducing perceived latency.

Gemini 3 Flash uses an internal chain-of-thought mechanism where the model breaks down complex problems into reasoning steps before generating final answers. While the reasoning process is not exposed by default, the model's training emphasizes step-by-step problem decomposition, enabling it to handle multi-step logic, math problems, and complex decision-making. This is particularly optimized for agentic workflows where intermediate reasoning must be reliable.

Gemini 3 Flash accepts a system prompt (or 'system instruction') that defines the model's behavior, tone, and constraints for a conversation. The system prompt is processed separately from user messages and influences all subsequent responses in the conversation without being repeated. This enables role-based customization (e.g., 'You are a Python expert', 'Respond in JSON only') that persists across multiple turns without token overhead.

Gemini 3 Flash supports batch API processing where multiple requests are submitted together and processed asynchronously, typically at a 50% cost reduction compared to real-time API calls. Batch requests are queued and processed during off-peak hours, with results delivered via webhook or polling. This is implemented via a separate batch endpoint that accepts JSONL-formatted request files and returns results in the same format.

Gemini 3 Flash includes built-in safety filters that detect and block harmful content (hate speech, violence, sexual content, etc.) before generation. The model returns safety ratings for each content category along with a block reason if content is filtered. Applications can configure safety thresholds per category (BLOCK_NONE, BLOCK_ONLY_HIGH, BLOCK_MEDIUM_AND_ABOVE, BLOCK_LOW_AND_ABOVE) to customize filtering strictness without retraining.