Gemini 3.1 Flash-Lite
A fast, low-cost multimodal reasoning model for high-volume tasks
Gemini 3.1 Flash-Lite is a natively multimodal reasoning model from Google DeepMind, published on 3 March 2026 as part of the Gemini 3 series. Built on Gemini 3 Pro and trained on Google TPUs, it targets high-volume, latency-sensitive work such as translation and classification with a 1M-token context window.
- Published 3 March 2026 as part of the Gemini 3 series
- Based on Gemini 3 Pro; architecture details sit on that model's card
- Accepts text, images, audio and video with a 1M-token context window
- Trained on Google TPUs using JAX and ML Pathways
- Reported output speed of 363 tokens per second as of March 2026
What it is
Gemini 3.1 Flash-Lite is a natively multimodal reasoning model in Google DeepMind's Gemini 3 series, published as a model card on 3 March 2026. It is built on Gemini 3 Pro and is designed to be cost-efficient and quick, aimed at high-volume, latency-sensitive work such as translation and classification. It accepts text, images, audio and video, with a token context window of up to 1M. Google DeepMind positions it for applications needing high throughput at low cost and low latency. Parameter counts and training data details are not given on its own card; readers are directed to the Gemini 3 Pro model card for architecture, training data and processing.
How it came about
The model card was published on 3 March 2026. Gemini 3.1 Flash-Lite is described as an addition to the Gemini 3 line, and it is based on Gemini 3 Pro. Its predecessor in the smaller tier is Gemini 2.5 Flash-Lite, which is the comparison point for the internal safety evaluations reported on the card. Model cards are updated over time as models are improved or revised.
Inside the model
Architecture and training data are documented on the Gemini 3 Pro model card, since this model is based on Gemini 3 Pro. Training used Google's Tensor Processing Units, which the company describes as designed for the large computations of language model training, with high-bandwidth memory and TPU Pods for scaling across devices. Training ran on JAX and ML Pathways. Inputs cover text strings, images, audio and video files, with a context window of up to 1M tokens. Google links the TPU efficiencies to its sustainability commitments.
What it can do
Google DeepMind reports March 2026 results across speed, reasoning, multimodal understanding, factuality, tool use, multilingual performance, coding and long context. Reported figures include 86.9% on GPQA Diamond, 76.8% on MMMU-Pro, 84.8% on Video-MMMU, 88.9% on MMMLU, 72.0% on LiveCodeBench, and 60.1% on MRCR v2 at 128k average, falling to 12.3% at 1M pointwise. Output speed is given as 363 tokens per second. The card compares these against several other models but does not rank them.
How to use it
The model is distributed through an application program interface for downstream providers, under relevant terms of use. No specific hardware or software is required. Access runs through AI Studio and the Gemini API under the Gemini API Additional Terms of Service, or through Vertex AI under the Google Cloud Platform Terms of Service. The company provides API instructions and a Vertex AI quickstart.
Safety and testing
Internal automated evaluations during development compared the model with Gemini 2.5 Flash-Lite. Google DeepMind says it outperforms that model on safety and tone while keeping unjustified refusals low, though image-to-text safety showed a -21.7% change and unjustified refusals -14.41%. Flagged content was manually reviewed, and the company says losses were mostly false positives or not egregious. Specialist red teams outside the development team ran manual testing; child safety launch thresholds were met, and no egregious concerns were found. Frontier safety relies on the assessment of Gemini 3.1 Pro with Deep Think, which reached no Critical Capability Levels.
Questions people ask about Gemini 3.1 Flash-Lite
What is Gemini 3.1 Flash-Lite for?
Google DeepMind describes it as a cost-efficient, fast model for high-volume, latency-sensitive work, giving translation and classification as examples. It is part of the Gemini 3 series of natively multimodal reasoning models and is based on Gemini 3 Pro.
What can it take as input?
It accepts text strings such as questions, prompts or documents to summarise, plus images, audio and video files. The token context window runs up to 1M. Output details beyond text responses are not specified on its own model card.
How do I get access to it?
It is offered through an API for downstream providers, with no required hardware or software. You can use it via AI Studio and the Gemini API under the Gemini API Additional Terms of Service, or via Vertex AI under the Google Cloud Platform Terms of Service.
How does it compare with the earlier Flash-Lite?
Google DeepMind's internal automated evaluations compare it with Gemini 2.5 Flash-Lite, reporting better overall safety and tone with low unjustified refusals. Some measures moved the other way, including image-to-text safety at -21.7% and unjustified refusals at -14.41%.
How large is the model?
No parameter count is given on the Gemini 3.1 Flash-Lite model card. The card points readers to the Gemini 3 Pro model card for architecture, training dataset and training data processing, since this model is based on Gemini 3 Pro.
Good reading and sources
There is no English Wikipedia article on Gemini 3.1 Flash-Lite yet, so the facts on this page come from the sources listed above: deepmind.google/models/model-cards/gemini-3-1-flash-lite/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.