Gemini 3 Flash Preview: Specs and Evaluation Steps
Review the Gemini 3 Flash Preview release, including its context window, modalities, and documented capabilities. Learn how to test this model effectively.
Verified Release Details
Google announced the launch of Gemini 3 Flash Preview on December 17, 2025. The model identifier is gemini-3-flash-preview. According to official documentation, the model supports text and image inputs with text output. It features a context window of 1,048,576 tokens and a maximum output limit of 65,536 tokens. The source notes that open weights are not confirmed for this release.
- Provider: Google
- Input Modalities: Text, Image
- Output Modality: Text
- Context Window: 1,048,576 tokens
- Max Output Tokens: 65,536
Documented Capabilities
The official changelog describes the model as delivering fast frontier-class performance. It states that the model rivals larger models at a fraction of the cost. Specific feature updates include upgraded visual and spatial reasoning, along with agentic coding capabilities. Additionally, the changelog records that support for the Computer Use tool was launched for both gemini-3-pro-preview and gemini-3-flash-preview. The alias gemini-flash-latest was switched to point to gemini-3-flash-preview.
- Upgraded visual and spatial reasoning
- Agentic coding capabilities
- Support for the Computer Use tool
- Alias update: gemini-flash-latest now maps to gemini-3-flash-preview
Evaluation Procedure
To assess the model's performance, engineers should design specific tests based on the documented features. First, create a dataset of complex images requiring spatial analysis. Measure the accuracy of the model's descriptions against ground truth data. Next, define a set of coding tasks that require agentic behavior. Validate the model's ability to execute multi-step code generation or modification. Compare the results against baseline expectations derived from your own internal standards. Do not assume cost savings or speed improvements without measuring them directly in your specific environment. The source confirms the existence of these features but does not provide benchmark scores or guaranteed operational metrics.
- Test visual reasoning with standardized image sets
- Measure coding task completion rates
- Compare outputs against internal baselines
- Validate Computer Use tool integration
Artificial Analysis benchmark review
Benchmark review updated on . These scores apply to the linked artifact and recorded methodology; they do not establish the performance of a different runtime or quantized model.
| Measurement | Recorded result | Measured |
|---|---|---|
| Artificial Analysis Intelligence Index (v4.1) ↗ | 27.9 index_points | 2026-08-09 |
| Artificial Analysis Intelligence Index (v4.2) ↗ | 20.8 index_points | 2026-09-05 |
| Artificial Analysis Intelligence Index (v4.3) ↗ | 17.9 index_points | 2026-09-09 |
Review the full benchmark provenance and source methodology →
Primary sources
Frequently asked questions
What is the exact date of the announcement?
The source indicates the event occurred on December 17, 2025. This is distinct from the current publication date of this article.
Are the model weights available for local installation?
The source records that open weights are not confirmed. It does not state whether local installation is possible or impossible, nor does it specify hosting requirements.