The fastest lightweight model for high-throughput agents
The fastest, lowest-cost 3.5 tier retains 1M context, native multimodal input, and adjustable thinking.
Overview
Gemini 3.5 Flash-Lite is Google DeepMind's lightweight multimodal reasoning model, generally available since July 21, 2026. Built on Gemini 3.1 Flash-Lite, it is the fastest and lowest-cost model in the 3.5 family, reaching 350 output tokens per second in Artificial Analysis measurements. It targets translation, classification, document parsing, and subagent workloads where concurrency and latency dominate.
It retains adjustable thinking levels, native multimodal input, structured outputs, and tool use. Across several agentic and coding evaluations it substantially improves on 3.1 Flash-Lite and sometimes exceeds Gemini 3 Flash. CrossModel exposes it as gemini/gemini-3.5-flash-lite.
Key capabilities
| Dimension | Detail |
|---|---|
| Context window | 1,048,576 tokens (about 1M) |
| Max output | 65,536 tokens (about 64K) |
| Input modalities | Text, image, audio, video |
| Output modalities | Text |
| Tools and controls | function calling, structured outputs, search grounding, code execution, adjustable thinking levels |
The default thinking level is
minimal; raise it for harder tasks. Standard pricing has no long-context surcharge, and cached input costs one tenth of fresh input. See live rates in the model catalog.
Agentic and coding performance
A lightweight tier that still handles multi-step engineering
Official July 2026 evaluation; Gemini results use high thinking and default sampling.
On Google's evaluations, 3.5 Flash-Lite scores 54.2% on SWE-Bench Pro, up from 38.3% for 3.1 Flash-Lite and close to GPT-5.4 mini's 54.4%. It reaches 54.0% on Terminal-bench 2.1 and 39.2% on MLE-Bench, showing that it can handle multi-step code changes, terminal execution, and ML experiments rather than only one-shot classification.
Its 74.0% OSWorld-Verified score is also above GPT-5.4 mini's 72.1%. For subagents that repeatedly inspect interfaces and call tools, that mix of reliability and low latency matters more than a single knowledge score.
Multimodal and long context
Tools improve chart synthesis; 128K retrieval steps forward
The 1M MRCR v2 value is pointwise at the full window; retrieval and chunking remain advisable.
On CharXiv chart synthesis it scores 74.5% without tools and 76.5% with search and code execution. GDM-MRCR v2 long-context retrieval reaches 72.2% at 128K, well ahead of 3.1 Flash-Lite's 60.1%, while the full 1M pointwise score is 21.3%. For full-window workloads, retrieval and chunking remain safer than expecting lossless recall from a packed context.
When to use it
- High-throughput document pipelines: extract, classify, translate, and structure receipts, contracts, tables, and email.
- Subagent execution: let a stronger coordinator split work while Flash-Lite runs searches, code, and tools in parallel.
- Low-latency interaction: support routing, input correction, moderation, and responsive UI assistance.
- Cost-sensitive batch work: data cleaning, summaries, tagging, and repository triage at scale.
Choose Gemini 3.6 Flash for a higher ceiling on complex planning and coding; choose Gemini 3.5 Flash-Lite when many well-scoped tasks need to finish quickly. CrossModel exposes it through an OpenAI-compatible /v1/chat/completions API. Current pricing is available in the model catalog.