Gemini 3 Flash Preview

Overview. Gemini 3 Flash Preview is Google's closed-weight, multimodal instruct model, accepting text, image, audio, and video input and returning text output. Released in December 2025, it carries a 1,048,576-token context window and sits in the fast, Flash-tier of the Gemini 3 line. Developers reach for it when they need quick multimodal Inference over long documents or mixed-media inputs without committing to a heavier flagship model.

Key strengths. Artificial Analysis benchmarks show a GPQA Diamond score of 0.81 and an MMMU-Pro score of 0.79, both strong results for a Flash-tier model and evidence of solid graduate-level reasoning and multimodal understanding. Reasoning effort is configurable from minimal to high (default medium), so a single model can flex between fast, lightweight responses and deeper multi-step reasoning on demand. It also supports vision, audio, and document inputs alongside tool calling and structured/JSON output.

When to use it. Use this model for interactive, latency-sensitive workloads that involve mixed media — images, audio clips, or long documents — where the million-token-plus context window matters. Its opt-in reasoning effort makes it a good fit for tasks that are usually simple but occasionally need a harder reasoning pass, without switching to a different model. It is well suited to complex analytical or research-style prompts, but it is an instruct model with reasoning as an add-on rather than a dedicated reasoning-first model.

Limitations. Google has not disclosed parameter count, architecture details (including whether it uses a mixture-of-experts design), or a training data cutoff for this checkpoint. Output is text-only — there is no native audio or image generation. Independent leaderboard standing (LMSys Arena) could not be confirmed in the latest research pass due to a rate-limited fetch, so its ranking against other frontier-class models remains unresolved.

Sovereignty note. Gemini 3 Flash Preview is built by Google, a US-based creator subject to CLOUD Act exposure, and served exclusively through Google's own proprietary API — there is no open-weight or self-hosted option.