TechPrufer
Get started
DeepMind
DeepMind logoDeepMind · Feature trail

Gemini Flash Model Series

21 posts · Oct 2025 – Sep 2026

What this trail covers

Gemini 3.1 Pro is released, offering enhanced core reasoning capabilities for complex problem-solving and agentic workflows. It achieves a verified score of 77.1% on the ARC-AGI-2 benchmark, more than doubling the reasoning performance of Gemini 3 Pro. The model is being rolled out across consumer and developer products, including the Gemini app, NotebookLM, Gemini API, Google AI Studio, Gemini CLI, Google Antigravity, Vertex AI, Gemini Enterprise, and Android Studio. Higher limits are available for Pro and Ultra plan users. The release is in preview for developers to validate updates and further advance agentic workflows.

All DeepMind trails
  1. Proactive cyber defense for governments and enterprisesIntroduces the Fairwind Program to provide early access to Gemini 3.8 Flash Cyber and the CodeMender harness for proactive cyber defense. This offering enables autonomous vulnerability discovery and automated patching, delivering verified, deployment-ready fixes in minutes. The program targets governments, critical infrastructure operators, and core technology platforms, emphasizing responsible use and operational standards. It aims to shrink the time between detecting a flaw and patching it, providing a security advantage against agentic-speed threats.
  2. Introducing Gemini 3.8 Flash and 3.8 Flash CyberIntroduces Gemini 3.8 Flash, an advancement over 3.7 Flash with improved performance on long-horizon coding (DeepSWE v1.1) and specialized reasoning tasks (Vals Finance Agent V2, Harvey's Legal Agent Benchmark, HLE-Verified). It achieves this through more diligent execution of reasoning steps and iterative tool calls. Also introduces Gemini 3.8 Flash Cyber, a cybersecurity model with frontier-level performance on vulnerability discovery (CyberGym, internal benchmark) and automated patching (CWE-Bench). Real-world impact is demonstrated through its use in securing Google's code (Chrome, Wiz, Cloud Vulnerability Research). Both models incorporate enhanced prompt injection robustness.
  3. Introducing Gemini 3.7 FlashGemini 3.7 Flash shows gains over 3.6 Flash in coding tasks (43.6% vs 34.4% in FrontierCode 1.1 Main, 65.3% vs 49.0% in DeepSWE v1.1), web development (Elo score of 1588 vs 1538 on WebDev Arena), and document processing (34.0% vs 22.0% on GDP.pdf, 30.4% vs 17.0% on AutomationBench). It offers improved multi-step planning and tool calls, and is available at an introductory price of $0.75/1M input tokens and $3.75/1M output tokens.
  4. Introducing Gemini 3.5 Flash CyberIntroduced Gemini 3.5 Flash Cyber, a specialized, lightweight cybersecurity model derived from Gemini 3.5 Flash. This model is fine-tuned for finding, validating, and patching software vulnerabilities, offering improved efficiency and cost-effectiveness compared to larger cybersecurity models. It is designed to be invoked multiple times by agents like CodeMender to analyze a large number of code paths and discover vulnerabilities. Benchmark results on CyberGym, Google's Big Sleep evaluation, and Chrome's production commit scanning pipeline demonstrate its competitive performance and ability to find more unique vulnerabilities than mainline Flash models. The model is being deployed in Google's internal codebases for real-world application.
  5. Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash CyberIntroduces Gemini 3.6 Flash with 17% fewer output tokens and improved performance in coding and knowledge work benchmarks (DeepSWE, MLE Bench, OSWorld-Verified, GDPval-AA v2). Releases Gemini 3.5 Flash-Lite, achieving 350 output tokens/s and outperforming 3.1 Flash-Lite in agentic tasks (Terminal-Bench 2.1, GDM-MRCR v2, GDPval-AA v2), and even surpassing 3 Flash in some coding and agentic evals (SWE-Bench Pro, OSWorld-Verified). Launches Gemini 3.5 Flash Cyber for cybersecurity, fine-tuned for vulnerability detection and remediation, integrated with the CodeMender agent and available via a limited pilot program. Enhances safety safeguards for 3.6 Flash in CBRN and cyber offense domains.
  7. AI in Indian Education: Atal Innovation Mission and Google launch ATL SaathiThe ATL Saathi application leverages Gemini 3.5 Flash to provide educators with a 24/7 planning and training assistant, transforming Tinkering Labs into AI-Augmented Discovery Labs. It offers streamlined onboarding and content curation using NotebookLM, advanced project generation with push & pull mentorship capabilities, and multilingual accessibility in 8 languages. The model generates grade-appropriate project ideas, detailed assembly instructions, wiring diagrams, and safety precautions.
  8. Introducing computer use in Gemini 3.5 FlashGemini 3.5 Flash now natively integrates 'computer use' capabilities, previously a standalone model. This enhancement allows developers to build custom agents that can interact with browser, mobile, and desktop environments, improving performance for long-horizon and enterprise automation tasks like continuous software testing. The post details targeted adversarial training for prompt injection risks and introduces two optional enterprise safeguard systems: explicit user confirmation for sensitive actions and automatic task stopping for indirect prompt injection.
  9. DiffusionGemma: 4x faster text generationIntroduces DiffusionGemma, an experimental 26B MoE model that utilizes text diffusion for up to 4x faster text generation on GPUs by processing entire blocks of text simultaneously. Key technical contributions include a novel diffusion head for speed, bi-directional attention enabling parallel token processing (256 tokens at a time), and iterative self-correction. It highlights trade-offs in output quality compared to autoregressive models and its suitability for low-concurrency, local inference scenarios. It also details integration with various serving and fine-tuning tools.
  10. Fluid, natural voice translation with Gemini 3.5 Live TranslateIntroduced Gemini 3.5 Live Translate, an audio model for live speech-to-speech translation. Key technical advancements include continuous speech generation that balances context and immediacy, automatic detection of 70+ languages, preservation of speaker intonation, pacing, and pitch, and noise robustness. The model is being integrated into Google Translate, Google Meet, and made available via the Gemini Live API for developers. SynthID watermarking is applied to all generated audio.
  11. Introducing Gemma 4 12B: a unified, encoder-free multimodal modelIntroduces Gemma 4 12B, a unified, encoder-free multimodal model. Key technical contributions include: replacing traditional multimodal encoders with a lightweight embedding module for vision and projecting raw audio signals directly into the LLM's dimensional space. This architecture reduces latency and memory usage, enabling local execution on consumer laptops. It also highlights advanced reasoning capabilities and the integration of Multi-Token Prediction (MTP) drafters.
  12. Gemini 3.5: frontier intelligence with actionThis post introduces Gemini 3.5 Flash, a new model in the Gemini Flash series. It highlights its performance on agentic and coding benchmarks (Terminal-Bench 2.1, GDPval-AA, MCP Atlas), multimodal understanding (CharXiv Reasoning), and output tokens per second (4x faster than other frontier models). It emphasizes its ability to handle long-horizon agentic tasks at scale, reducing development time and cost. The post also mentions its integration with the Antigravity harness for deploying collaborative subagents and its ability to generate richer web UIs and graphics. It also touches on the development of Gemini 3.5 Pro and the safety frameworks used.
  13. Gemma 4: Byte for byte, the most capable open modelsIntroduced Gemma 4, a new family of open models built on Gemini 3 technology. This release includes four sizes (E2B, E4B, 26B MoE, 31B Dense) optimized for advanced reasoning and agentic workflows. Key technical features include native support for function-calling, structured JSON output, system instructions, code generation, multimodal processing (vision and audio), and extended context windows (up to 256K). The models are designed for efficient deployment across diverse hardware, from edge devices to cloud, and are released under an Apache 2.0 license.
  14. Gemini 3.1 Flash Live: Making audio AI more natural and reliableIntroduced Gemini 3.1 Flash Live, a new audio and voice model that significantly enhances real-time dialogue capabilities. This model achieves higher quality, speed, and natural rhythm for voice-first AI. It demonstrates improved performance on benchmarks like ComplexFuncBench Audio (90.8%) and Scale AI's Audio MultiChallenge (36.1% with 'thinking' on), indicating robust reasoning and complex instruction following. Enhanced tonal understanding allows for more natural dialogue and dynamic response adjustment to user emotions. The model is inherently multilingual, supporting global expansion of services. Audio output is watermarked with SynthID for AI-generated content detection.
  15. Gemini 3.1 Flash-Lite: Built for intelligence at scaleIntroduces Gemini 3.1 Flash-Lite, a new model in the Gemini 3 series, focusing on cost-efficiency and speed for high-volume developer workloads. It offers a 2.5X faster Time to First Answer Token and 45% increase in output speed compared to 2.5 Flash, with an Elo score of 1432 on Arena.ai. It includes 'thinking levels' for adaptive reasoning, making it suitable for tasks like translation, content moderation, UI generation, and simulations. Early testers highlighted its efficiency, reasoning capabilities, and adherence to instructions.
  16. Gemini 3.1 Pro: A smarter model for your most complex tasksThis post announces the release of Gemini 3.1 Pro, an upgraded core intelligence for Gemini models. It highlights significant improvements in reasoning capabilities, evidenced by a 77.1% score on the ARC-AGI-2 benchmark, which is more than double the performance of Gemini 3 Pro. The post details the availability of Gemini 3.1 Pro across various consumer and developer platforms, including the Gemini API, Google AI Studio, Gemini CLI, Google Antigravity, Vertex AI, Gemini Enterprise, and Android Studio, with a focus on its application in complex problem-solving and agentic workflows. It also mentions higher usage limits for Pro and Ultra plan users and its availability on NotebookLM.
  17. Gemini 3 Flash: frontier intelligence built for speedThis post announces the release of Gemini 3 Flash, a new model in the Gemini 3 family. It highlights its speed, efficiency, and cost-effectiveness compared to previous models (Gemini 2.5 Pro and Gemini 3 Pro). Key technical details include its performance on reasoning benchmarks (GPQA Diamond, Humanity's Last Exam, MMMU Pro), its ability to modulate thinking for efficiency (30% fewer tokens on average than 2.5 Pro), and its speed advantage (3x faster than 2.5 Pro). For developers, it emphasizes its Pro-grade coding performance with low latency, achieving a score of 78% on SWE-bench Verified. It is positioned as ideal for agentic coding, production-ready systems, and responsive interactive applications, enabling complex video analysis, data extraction, and visual Q&A. The post also details its rollout to consumers via the Gemini app and AI Mode in Search, and to enterprises via Vertex AI and Gemini Enterprise.
  18. A new era of intelligence with Gemini 3This post introduces Gemini 3, the latest iteration of the Gemini model series. It highlights significant advancements in reasoning, multimodal understanding, and agentic capabilities. Key technical contributions include: state-of-the-art performance on various benchmarks (LMArena, Humanity's Last Exam, GPQA Diamond, MathArena Apex, MMMU-Pro, Video-MMMU, SimpleQA Verified, WebDev Arena, Terminal-Bench 2.0, SWE-bench Verified, Vending-Bench 2), improved zero-shot generation for UI rendering, enhanced agentic coding with direct editor/terminal/browser access via Google Antigravity, and improved long-horizon planning. Gemini 3 Deep Think mode is also introduced as an enhanced reasoning engine.
  19. Start building with Gemini 3This post introduces Gemini 3 Pro, highlighting its advancements in agentic coding, multimodal understanding (visual, spatial, and video reasoning), and its integration into developer tools like Google Antigravity, Gemini CLI, and Google AI Studio. It details new API parameters for controlling media resolution and a 'thinking level' for enhanced reasoning. The post emphasizes the model's ability to translate high-level ideas into functional applications with single prompts and its improved performance on benchmarks like MMMU-Pro.
  20. MedGemma: Our most capable open models for health AI developmentIntroduced MedGemma, a new series of open models derived from Gemini 3 Flash, specifically tailored for health AI development. These models are optimized for high-volume workloads, offering improved speed and quality for complex health-related tasks. MedGemma models are designed to support a broad spectrum of health AI applications and are available in multiple sizes for flexible integration.