Gemini Nano

Gemini Nano is the smallest and most efficient variant of Google's Gemini family of multimodal large language models, designed to run directly on phones and other edge hardware instead of in cloud data centers.^[1] Google introduced it on December 6, 2023 as the on-device tier of Gemini 1.0, ships it in two sizes, Nano-1 with 1.8 billion parameters and Nano-2 with 3.25 billion parameters, both distilled from larger Gemini models and quantized to 4 bits, and first deployed it on the Pixel 8 Pro to power Summarize in Recorder and Smart Reply in Gboard.^[1][2][3] Because the model runs locally, it works offline and keeps user data on the device: Google says Gemini Nano on the Pixel 8 Pro "offers several advantages by design, helping prevent sensitive data from leaving the phone, as well as offering the ability to use features without a network connection."^[3]

Gemini Nano was announced by Google CEO Sundar Pichai and Google DeepMind CEO Demis Hassabis alongside the larger Gemini Pro and Gemini Ultra models, and first shipped on the Pixel 8 Pro in the December 2023 Pixel Feature Drop.^[1][3] Nano powers on-device features such as summarization in the Recorder app, Smart Reply in Gboard, Magic Compose in Google Messages, Pixel Screenshots search, and image descriptions in TalkBack, while keeping user data on the device.^[3][11]

The original Gemini technical report (arXiv:2312.11805) describes Nano as a pair of small models distilled from larger Gemini variants and quantized to 4 bits, and frames the series as "the best-in-class small language models for a wide variety of tasks, such as summarization and reading comprehension, which power our next generation on-device experiences."^[2] Gemini Nano-1 has 1.8 billion parameters and targets low-memory devices, while Gemini Nano-2 has 3.25 billion parameters and targets higher-memory devices.^[2][33] Google has since expanded Nano beyond Pixel through the Android AICore system service and the Google AI Edge SDK, brought it into desktop Chrome through the built-in Prompt API, and added a multimodal Nano with image and audio understanding on the Pixel 9 series.^[7][9][17] With Apple's 2024 release of Apple Intelligence, Gemini Nano became the most prominent counterpart to Apple's on-device AI foundation models in the smartphone market.^[21]

When was Gemini Nano released?

Google announced the Gemini family in a blog post by Sundar Pichai and Google DeepMind CEO Demis Hassabis on December 6, 2023.^[1][6] Pichai described Gemini as Google's "most capable and general model yet," and the post said it was "built from the ground up to be multimodal," handling text, code, audio, images, and video together rather than stitching modalities together later.^[1] The first generation, Gemini 1.0, was released in three sizes: Ultra for highly complex tasks, Pro for a balance of capability and scale, and Nano for on-device tasks.^[1]

The accompanying technical report, "Gemini: A Family of Highly Capable Multimodal Models," appeared on arXiv as paper 2312.11805 on December 19, 2023, with later revisions through 2024 and 2025.^[2] The report describes Gemini Nano as two small models, Nano-1 with 1.8 billion parameters and Nano-2 with 3.25 billion parameters, both trained by distilling from larger Gemini models and quantized to 4 bits.^[2] The paper frames Nano as the variant intended for memory-constrained, on-device deployment, in contrast to Pro and Ultra, which were initially served from Google's data centers.^[2]

The Pixel 8 Pro was the first device to ship Gemini Nano. Google's December 2023 Pixel Feature Drop, also published on December 6, 2023, announced that Nano was running on the phone and powering two production features: Summarize in Recorder, which generated bullet-point summaries of voice recordings without a network connection, and Smart Reply in Gboard, which produced contextual responses for messaging apps including WhatsApp, Line, and KakaoTalk.^[3] Both ran locally on the Tensor G3 chip.^[3][4] Google's blog called the Pixel 8 Pro "the first smartphone with AI built in," referring to the system-level integration of an on-device large language model, and described Gemini Nano as "our most efficient model built for on-device tasks."^[3]

Behind the scenes, Google introduced Android AICore the same day as a new Android 14 system service that provides apps with access to Gemini Nano while managing the model, runtime, and safety features.^[5] AICore launched with support for Pixel 8 Pro and was designed from the start to accept neural-network co-processors from Qualcomm, Samsung S.LSI, and MediaTek, signalling that Google intended Nano to be a portable component across the Android ecosystem rather than a Pixel-exclusive.^[5]

How does Gemini Nano fit in the Gemini family?

Gemini Nano sits inside a wider model family that has grown across multiple generations.

Nano is the only branch explicitly designed for on-device use. The cloud Flash models are smaller and cheaper than Pro or Ultra but still run in Google's data centers. Nano carries the Gemini brand into phones, watches, and browsers, and Google has shipped several updated Nano versions since 2023.

How is Gemini Nano built?

Gemini Nano shares the broader Gemini architecture, a decoder-only Transformer trained on text, code, images, and audio.^[2] Two design choices distinguish Nano from the Pro and Ultra variants: distillation and aggressive quantization.

Distillation from larger Gemini models

Knowledge distillation means the Nano models are trained to imitate the outputs of larger Gemini models rather than learning purely from raw training data.^[2] The paper gives no detail on the distillation recipe beyond saying Nano "is trained by distilling from larger Gemini models," and credits "additional advancements in distillation and training algorithms" for the series' results.^[2] The Gemini 1.0 report frames the Nano series as "the best-in-class small language models for a wide variety of tasks, such as summarization and reading comprehension," and pairs distillation with 4-bit quantization for deployment.^[2]

4-bit quantization

Quantization to 4 bits per weight roughly quarters the storage and memory bandwidth required compared to a 16-bit floating point representation. By simple arithmetic from the parameter counts, 4-bit weights alone come to roughly 0.9 gigabytes for Nano-1 and 1.6 gigabytes for Nano-2, before activations, cache, and any LoRA (Low-Rank Adaptation) adapters loaded on top; no official memory figures have been published. Google has not published the details of the quantization scheme; the 4-bit figure comes from the Gemini report, while Android documentation says AICore runs Nano on the Tensor TPU and on NPUs in partner silicon.^[2][5]

Compared to Apple's on-device foundation model, which uses a mixed 2-bit and 4-bit configuration averaging 3.7 bits per weight, Nano's uniform 4 bits per weight is slightly less aggressive.^[2][22] Phi-3 Mini from Microsoft can also be quantized to 4 bits and occupies roughly 1.8 gigabytes in that form, slotting between Nano-1 and Nano-2 on a memory-versus-parameter axis.^[23]

Is Gemini Nano open source?

Google has kept Nano as a closed-weights model. There are no public weight downloads, and developers interact with Nano only through Google's APIs on Android, Chrome, and partner devices. The related Gemma family is Google's open-weights line of small Transformer models, but it is a separate brand with its own checkpoints; Nano remains the proprietary on-device model that Google ships through its own surfaces. The two lines share research lineage and tooling. An April 2026 Android Developers Blog post said Gemma 4 is the foundation for the next generation, Gemini Nano 4, and that Gemma 4 code would work on Nano 4 devices, which began shipping in August 2026; the Nano weights Google ships through AICore are not offered as downloads.^[37][38]

Multimodal Nano

The Gemini 1.0 report evaluated Nano-1 and Nano-2 on image understanding and Nano-1 on speech recognition, but the first Nano features shipped on the Pixel 8 Pro worked on text (Recorder transcripts and Gboard replies).^[2][3] At Google I/O in May 2024, Google previewed "Gemini Nano with Multimodality," a variant that adds image and audio understanding in addition to text.^[16] The multimodal Nano shipped on the Pixel 9 series in August 2024 and is "significantly larger than the previous one," roughly twice the size of Nano 1.0 and described by Google as "three times more capable and sophisticated" than the version that shipped on the Pixel 8 Pro.^[17] Google told reporters the upgraded model needs new NPUs and quite a bit of RAM; the Pixel 9 series ships with 12 to 16 gigabytes, part of which Android Authority says is reserved for AI models such as Nano.^[11][16]

What can Gemini Nano do?

Gemini Nano covers a narrower range of tasks than its cloud siblings, but it handles the everyday workloads that benefit most from low latency and on-device privacy. The model handles text generation and rewriting (including tone shifts used by Magic Compose in Google Messages), summarization of long inputs such as Recorder audio transcripts and chat threads, smart reply and short-message proofreading, and, on Pixel 9 and later, multimodal text-image-audio understanding through the Nano with Multimodality variant.^[11][14][17] The same multimodal model powers richer transcript summaries and detailed image descriptions for accessibility tools. Sensor- and context-aware features combine Nano output with on-device signals such as call audio, screenshots, or app context.

Pixel Screenshots, introduced with the Pixel 9 family in August 2024, is a representative case. It lets a user search saved screenshots in plain English and pulls out details like flight numbers, Wi-Fi passwords, or recipe ingredients.^[20] Android Authority notes that because the processing takes place offline, results are near-immediate even with thousands of screenshots in the gallery.^[11]

Call Notes, also on Pixel 9 and later, takes the audio of a phone call and produces a private on-device summary with the key details; both parties on the call are notified when the feature is active.^[20] In Pixel Weather, Gemini Nano generates a short custom weather report, which Android Authority says can take UV index and air quality into account.^[11][20] Pixel Studio, launched alongside these features, is not a Nano feature: Google says it is powered by an on-device diffusion model running on Tensor G4 together with its cloud Imagen 3 model.^[20]

Which devices and platforms run Gemini Nano?

Google has expanded Gemini Nano well beyond the original Pixel 8 Pro launch.

Samsung's Galaxy S24 series was the first non-Pixel phone to run Gemini Nano, bundled through AICore and powering on-device Magic Compose in Google Messages, while Samsung's Notes, Voice Recorder, and Keyboard apps use cloud-served Gemini Pro for summarization.^[12][13] The same AICore plumbing has since shipped on Motorola and Xiaomi flagships, with Tensor, Qualcomm Snapdragon, and MediaTek Dimensity NPUs as the supported acceleration paths.^[5][7]

The Pixel 9a, released in 2025, uses a smaller "Nano 1.0 XXS" variant; Google reduced the model because the device ships with only 8 gigabytes of RAM compared to the 12 to 16 gigabytes on the rest of the Pixel 9 series.^[18] Nano XXS is text-only, does not run in the background, and the Pixel 9a therefore loses several multimodal AI features available on the Pixel 9, including Pixel Screenshots and Call Notes.^[18]

How do developers access Gemini Nano?

The layer between user-facing apps and Gemini Nano on Android is built around two pieces. AICore is an Android system service introduced with Android 14. It manages the Nano model weights, runs inference on the Tensor TPU or partner NPU, applies safety filters, and updates the model out of band so apps do not have to bundle large weights themselves.^[5][7] Google describes AICore as "private by design": the service has no direct internet access, isolates each request, and follows Android's Private Compute Core architecture, introduced at Google I/O 2021.^[5][8]

The Google AI Edge SDK is the developer-facing API for AICore. The first SDK package, com.google.ai.edge.aicore:aicore:0.0.1-exp01, opened experimental access on October 1, 2024, initially limited to text-to-text prompts on Pixel 9 series devices.^[9] The SDK lets a developer set parameters such as temperature, top K, candidate count, and max output tokens, and stream tokens as they are produced.^[9] The wider Google AI Edge effort also covers LiteRT (formerly TensorFlow Lite (LiteRT)) and a broader runtime called LiteRT-LM aimed at small language models on edge devices, with Google describing LiteRT-LM as "the production-ready inference framework that has been powering some of the widest deployments of Gemini Nano across Google products to date" and a "battle-tested engine" for Nano in Chrome, Chromebook Plus, and the Pixel Watch.^[28][34]

LoRA adapters and safety pipeline

AICore exposes LoRA (Low-Rank Adaptation) fine-tuning on top of the base Nano weights, so feature owners can specialize the model for a single task without retraining or duplicating the full network.^[5][8] Google uses the same mechanism internally. ML Kit's summarization, proofreading, and image-description APIs each ship with a small API-specific LoRA adapter trained on representative data; the adapter loads on demand once the base Nano model is on the device.^[10] Google reports that feature-specific LoRA tuning lifts the summarization quality benchmark from 77.2 to 92.1 and image description from 86.9 to 92.3 on its internal raters.^[10]

LoRA tuning also doubles as the integration point for safety controls. The LoRA blocks are trained against app-specific safety data alongside the task data, and AICore runs additional input and output safety classifiers around each call so that the same base model can satisfy different safety standards for different surfaces, from Recorder summaries to keyboard Smart Reply.^[8][10]

ML Kit GenAI APIs

ML Kit GenAI APIs sit one level above the AI Edge SDK and give Android developers ready-made entry points. The first four, Summarization, Proofreading, Rewriting, and Image Description, were announced at Google I/O in May 2025;^[10][35] by late 2026 Google's documentation also listed a Prompt API (direct model access), Speech Recognition, speech translation, text to speech, and an embedding API for semantic search.^[7] Unlike the earlier AI Edge SDK preview, ML Kit GenAI runs on a wider range of Android phones, supports image inputs for the Image Description API, and is intended for production use on devices with optimized MediaTek Dimensity, Qualcomm Snapdragon, or Google Tensor hardware through AICore.^[10][35]

How does the Chrome Prompt API use Gemini Nano?

In Chrome, Google added a built-in AI stack starting in 2024 with experimental flags and began shipping parts of it in stable with Chrome 138 in 2025.^[29][30] The Prompt API exposes Gemini Nano through a JavaScript LanguageModel global object, with promise-based and streaming interfaces for text generation; Google states that "with the Prompt API, you can send natural language requests to Gemini Nano in Chrome" and that "no data is sent to Google or any third party when using the model."^[29] Key methods include LanguageModel.availability(), LanguageModel.create() for instantiating a session, and session.prompt() or session.promptStreaming() for synchronous and streaming responses.^[29]

Browser support requires Windows 10 or 11, macOS 13 and later, recent Linux, or ChromeOS on Chromebook Plus, plus at least 22 gigabytes of free space on the Chrome profile volume (the model itself is significantly smaller).^[29] Hardware-wise, Chrome requires a GPU with more than 4 gigabytes of VRAM or a CPU with at least 16 gigabytes of RAM and four or more cores, with audio input restricted to GPU paths.^[29] The network is only needed for the one-time model download: Google notes that "subsequent use of the model does not require a network connection."^[29] Earlier reporting referred to the namespace as window.ai, but the released API uses LanguageModel directly and ships in Chrome rather than as a third-party shim.^[29]

The Prompt API shipped in Chrome 138 for Chrome Extensions and, according to Google's API status table, in Chrome 148 for web pages.^[29][30] Three sibling APIs reached stable in Chrome 138: the Summarizer API, the Translator API, and the Language Detector API.^[30] As of May 2025, the Writer and Rewriter APIs were in origin trials, the Proofreader API was headed to Chrome 139 Canary for local experimentation by Early Preview Program participants, and a multimodal variant of the Prompt API that accepts audio and images was available to those participants.^[30] The Prompt API accepts English, Japanese, Spanish, German, and French, with more languages in development.^[29]

What is Gemini Nano used for?

Production features using Gemini Nano on Android and Chrome include Summarize in Recorder (offline bullet summaries of voice recordings), Smart Reply in Gboard (contextual chips for WhatsApp, Line, KakaoTalk), Magic Compose in Google Messages (tone rewriting with formal, excited, chill, and Shakespearean styles operating on the last 20 messages), Pixel Screenshots (natural-language search through saved screenshots), Call Notes (call transcription and summary), Pixel Weather AI Reports, and TalkBack image descriptions.^{[3][11][12][20]} On Pixel 10, Google says Tensor G5 with Gemini Nano enables on-device features such as Magic Cue, Voice Translate, Call Notes with actions, and Personal Journal.^[25][26]

Google also says Tensor G5 with Gemini Nano enhances Scam Detection and natural voice editing in Gboard.^[26] Voice Translate runs a generative model in parallel with a classical audio model to translate a caller while preserving their voice, and Pro Res Zoom, which extends zoom to 100x, is a separate diffusion model of nearly 1 billion parameters rather than Gemini Nano.^[39]

Pixel Watch 4 and Pixel Buds 2a extend the Gemini assistant (not specifically the on-device Nano model) into wearables. Raise to Talk on Pixel Watch 4 starts a Gemini session by lifting the wrist, while Pixel Buds 2a are the first A-series buds to feature Gemini, which can summarize or respond to messages without the user checking the phone.^[25][27] Wear OS 6 began rolling Gemini out to Pixel Watch, Samsung, OnePlus, OPPO, and Xiaomi wearables through 2025.^[27]

In Chrome, the Prompt API lets web pages run summarization, translation, and generation directly in the browser. Third-party Android apps build on the same model through ML Kit GenAI, the AI Edge SDK, or AICore, paying nothing per call and keeping user data on the device.^[10][29]

How fast is Gemini Nano and how does it score?

Google has not published a full benchmark suite for the shipping Nano models, but several public numbers give a sense of the range. The Gemini 1.0 paper reports MMLU and BoolQ scores for Nano-1 and Nano-2 well below Pro and Ultra but competitive with other small Transformer models of similar size.^[2] In particular, Gemini Nano-1 scored 45.9% on MMLU (5-shot) and 71.6% on BoolQ in the launch paper (0.64 and 0.81 of Gemini Pro's scores), while Nano-2 scored 55.8% and 79.3% (0.78 and 0.90 of Pro).^[2] Phi-3 Mini, released roughly five months later, is reported to score 68.8% on the same MMLU benchmark, highlighting that small-model quality moved quickly in the year after Nano shipped.^[23]

Google's October 2024 AI Edge SDK post notes that the experimental on-device Nano (described as Nano 2) scored 56% on MMLU compared with 46% for the earlier version, 23% on math compared with 14%, 90% on paraphrasing compared with 44%, and 82% on smart reply compared with 44%.^[9] These are internal task evaluations, not standardized public benchmarks, but they show the trajectory of the on-device model since launch.

Google's ML Kit team reports that on a Pixel 9 Pro, Nano processes input at about 510 tokens per second and generates output at about 11 tokens per second.^[10] Tensor G5, the Pixel 10 chip manufactured on a TSMC 3 nanometer process, runs the newest Gemini Nano model approximately 2.6 times faster and 2 times more efficiently than the Tensor G4 inside the Pixel 9, and, according to 9to5Google, expands the on-device context window from 12,000 to 32,000 tokens.^[26][39] The Tensor G4 itself reuses the "rio" Edge TPU codename from the G3 with the same clock speed, which suggests that Nano gains between the Pixel 8 Pro and Pixel 9 owed more to the model and memory than to a new accelerator.^[31]

Memory footprint is dominated by the 4-bit weights (roughly 0.9 gigabytes for Nano-1 and 1.6 gigabytes for Nano-2 by arithmetic from the parameter counts), plus working memory for the KV cache.^[2] Longer Recorder summaries arrived with the larger multimodal model on the Pixel 9 series: Android Authority reported that the Pixel 8 series could summarize around fifteen minutes of audio, against more than half an hour on the Pixel 9.^[11][17]

Quality on hard tasks lags behind cloud-served Gemini Pro, Gemini Flash, and frontier models from other vendors. Features have improved with subsequent Nano updates and tighter prompt tuning, but the on-device model is best understood as a fast assistant for short, well-scoped tasks rather than a frontier reasoning system.

Google's internal benchmarks released alongside the AI Edge SDK preview showed clear gains as the on-device Nano matured. The developer build (the paper's Nano-2) scored 56% on MMLU against 46% for the smaller Nano-1, both Gemini 1.0 models, while paraphrasing rose from 44% to 90% and smart reply rose from 44% to 82%.^[9] Math also improved from 14% to 23%, although the model remains weak compared to Phi-3 Mini's 68.8% MMLU performance and Apple's larger on-device model's reported lead over Phi-3 Mini and Mistral 7B on human evaluation.^[22][23] The story across vendors is that the same parameter budget is delivering more useful capability every quarter, driven by data curation, distillation improvements, and feature-specific LoRA adapters rather than larger weights.^[10][22]

How does Gemini Nano compare with other on-device language models?

Gemini Nano is one of several small language model systems positioned for on-device use. The table compares Nano with the most prominent peers, using parameter counts disclosed in papers, blog posts, or model cards.

Gemini Nano is unusual on this list for two reasons. It ships pre-installed on consumer phones through a system service rather than as a hobbyist download, and it is gated behind Google's API rather than released as open weights. LLaMA and Phi-3 are open-weights models that rely on third-party runtimes such as llama.cpp or Ollama to run on phones, while Apple Intelligence is a parallel system-level integration locked to Apple silicon. GPT-4o mini is a cloud-hosted OpenAI model included in the table only to clarify it is not a peer of Nano in the on-device sense.

Microsoft's Phi-3 Mini, with 3.8 billion parameters and a similar 4-bit deployment profile, posts higher absolute MMLU scores than the Gemini Nano numbers in the original 1.0 paper but does not have an equivalent system-level integration in any consumer operating system.^[23] In practice, that distinction is what makes Nano commercially significant: a Pixel user does not pick Nano off a model hub, the operating system runs it on their behalf, much like Apple Intelligence on iOS.

How does Gemini Nano differ from Apple Intelligence?

Apple's response to Gemini Nano arrived at WWDC in June 2024 as Apple Intelligence.^[21][22] Both systems target similar features (writing tools, summaries, smart replies, image descriptions, on-device assistants), but they differ on hardware, model layout, and trust model.

Apple's machine-learning research paper claims that its 3 billion parameter on-device model outperforms Phi-3 Mini, Mistral 7B, Gemma 7B, and Llama 3 8B on its human evaluation suite, and Apple reports a generation rate of about 30 tokens per second on the iPhone 15 Pro with a time-to-first-token latency of about 0.6 milliseconds per prompt token.^[22] Both systems use small task-specific LoRA adapters on top of a shared base model.^[5][22]

The two systems represent different bets about what "system-level AI" looks like. Google's bet is that Nano can be a portable layer running on Pixel, partner Android phones, and desktop Chrome, with cloud Gemini available when the on-device model is not enough. Apple's bet is a tighter pair of one on-device model and one private-cloud model, both restricted to Apple Silicon, with third-party cloud models such as ChatGPT plugged in only on user request.

What are the limitations of Gemini Nano?

Gemini Nano is constrained by both its size and its deployment context. It is much smaller than cloud Gemini models, so it is weaker on hard reasoning, long-context tasks, and complex coding. The Gemini paper's own benchmarks show Nano-2 well behind Gemini Pro on MMLU, GSM8K, and similar evaluations.^[2] For tasks like explaining a legal contract or refactoring a large codebase, the cloud-served Gemini models and competing frontier systems are still the right tools.

4-bit quantization gives up some quality compared to higher-precision weights, in exchange for faster inference and lower memory. Battery, thermal, and memory constraints also limit how aggressively apps can use Nano. Google's documentation notes on-device constraints such as token limits and device hardware limitations, and suggests hybrid inference that falls back to cloud Gemini models when Nano is unavailable.^[7]

Device support is uneven. Many Nano features are exclusive to Pixel, with Samsung, Motorola, and Xiaomi getting subsets through partnerships.^[11][12] The Pixel 9a uses a smaller "Nano XXS" variant with only text input and no background execution, and older or lower-tier Android phones do not support Nano at all.^[18] Google's newer Gemini Intelligence layer, as of 2026, requires 12 gigabytes of RAM, a flagship chip, and a Gemini Nano v3 or higher runtime, leaving devices below that bar with reduced features.^[32]

Chrome's Prompt API requires at least 22 gigabytes of free disk space for the model download, and is not yet supported on mobile or on ChromeOS devices other than Chromebook Plus.^[29] Nano is also closed source: there are no public weights, and developers cannot fine-tune the base model directly. Google supports LoRA adapters on top of Nano through AICore for some customization, but the base model remains under Google's control.^[7][8]

The original Pixel 8 Pro launch features were narrow: Smart Reply started as a developer preview for WhatsApp, Line, and KakaoTalk, and Google limited the launch features to English (Smart Reply to the US English keyboard).^[3]

Recent developments

In January 2024, the Galaxy S24 series brought Gemini Nano to its first non-Pixel phone, alongside Circle to Search and Magic Compose.^[12][13] In March 2024 Google reversed an earlier statement and said the 8 GB Pixel 8 would get Gemini Nano as a developer preview with the June 2024 Feature Drop; the Pixel 8a launched in May 2024 with the same limited Nano features.^[15][16]

Google I/O 2024 previewed Gemini Nano with Multimodality, the variant that adds image and audio understanding alongside text.^[16] The Pixel 9 series, launched in August 2024, became the first phone to ship the multimodal Nano. 9to5Google reported the model is roughly twice the size of the original and substantially more capable on Recorder summaries and image descriptions.^[17] The same release added Pixel Screenshots, Pixel Studio, richer TalkBack image descriptions, and Call Notes.^[20]

On October 1, 2024, Google opened experimental access to Gemini Nano for all Android developers through the AI Edge SDK and AICore, initially limited to Pixel 9 devices for text-to-text prompts.^[9] In May 2025, ML Kit GenAI APIs added higher-level entry points for summarization, proofreading, rewriting, and image description, with prompt and speech APIs following later.^[7][10] Google I/O 2025 included a session titled "Gemini Nano on Android: building with on-device gen AI," framing Nano as a stable platform for third-party apps rather than a Pixel-only experiment.^[24]

In August 2025, the Pixel 10 series launched on the Tensor G5 with Nano behind features such as Magic Cue, Voice Translate, and Call Notes with actions, plus on-device portions of the photo-to-video feature in the Gemini app.^[25][26] The Tensor G5 expanded the on-device context window from 12,000 to 32,000 tokens and, by Google's figures, ran the newest Nano model 2.6 times faster and 2 times more efficiently than Tensor G4.^[26][39] The Pixel Watch 4 added Raise to Talk for Gemini, and Pixel Buds 2a added Gemini message summaries and replies.^[25][27]

An April 2, 2026 Android Developers Blog post announcing Gemma 4 in the AICore Developer Preview said the Gemma 4 models are "the foundation for the next generation of Gemini Nano," and that code written for Gemma 4 "will automatically work on Gemini Nano 4-enabled devices."^[37] The first Nano 4 devices, the Pixel 11 series and Samsung's Galaxy Z Flip8, Z Fold8, and Z Fold8 Ultra, launched in August 2026, when 9to5Google reported that Google's Gemini Intelligence features require Nano 3 or later.^[38] Chrome's built-in AI APIs reached stable with Chrome 138 in 2025, when the Prompt API shipped for Chrome Extensions; Google lists the web version of the Prompt API for Chrome 148, exposing Nano through the LanguageModel JavaScript object on Windows, macOS, Linux, and Chromebook Plus.^[29][30]

Strategic significance

Gemini Nano is the on-device anchor of Google's AI strategy. The cloud-served Pro, Flash, and Ultra models compete directly with OpenAI's GPT line and Anthropic's Claude in the API and chatbot markets, but Nano competes in a different fight: which AI assistant is built into the operating system a user picks up first thing in the morning. By shipping Nano on Pixel, partner Android phones, Pixel Watch, Chromebook Plus, and desktop Chrome, Google offers a system-level AI layer no pure API vendor can match.^[29][34]

Nano is also the most direct counterweight to Apple Intelligence. Both companies have settled on a similar pattern: a small on-device model for fast, private operations and a larger cloud model for harder work. The differences between them, on hardware, on disclosure, and on third-party integration, will shape how the phone industry thinks about AI for years.^[21][22]

For developers, Nano expands the surface area of generative AI. A prompt through the AI Edge SDK costs nothing per call, runs offline, and keeps user data on the device, which makes it attractive for features hard to justify with a cloud API: real-time keyboard help, sensitive document summaries, accessibility tools, and long-running background work.^[9][10] The trade-off is that the model is smaller, the API is gated to specific devices and chipsets, and Google sets the rules for what runs through AICore.^[7][8] Nano shifts some on-device AI design decisions from app developers to the platform vendor, much as camera pipelines and notification systems have done on smartphones. Generative AI is moving from a service users go to into a feature ambient in the phone, the watch, and the browser, and Gemini Nano is one of the clearest early examples of that shift in production.^[25][27][29]

For the browser ecosystem, the Chrome Prompt API turns Gemini Nano into the first widely available on-device LLM exposed through a standard web platform interface.^[29] That changes what a web developer can ship without infrastructure costs or privacy disclosures: a translation widget, a summarizer for long articles, a structured-output JSON generator for an internal tool, or a generative search box on a personal blog can all run entirely on the user's machine.^[29][30] The chief constraints are the 22 gigabyte disk requirement for the model and the relatively narrow set of supported operating systems, which Google has confirmed will expand over time.^[29] The combination of Android AICore on phones and Chrome's Prompt API in the browser gives Google two reference surfaces that no third-party LLM vendor can replicate without a comparable platform footprint.

Continue reading

• Gemini (language model)Gemini is a family of natively multimodal large language models developed by Google DeepMind, first announced on December 6, 2023, that can reason across text, images, audio, video, and code within a single…Mentioned 118 times above
• Gemini (app)The Gemini app is Google's consumer AI assistant, a chat application available on the web at gemini.google.com and as native Android, iOS, macOS and Windows appsMentioned 118 times above
• LLoRA (Low-Rank Adaptation)Low-Rank Adaptation, usually abbreviated LoRA, is a parameter-efficient fine-tuning method for adapting a pre-trained model.Mentioned 10 times above
• Phi-3Mentioned 8 times above
• QQuantizationQuantization is the representation of values from a large or continuous set by values from a smaller, usually finite set.Mentioned 8 times above
• MMLUMMLU, short for Massive Multitask Language Understanding, is an English-language AI benchmark for evaluating broad academic and professional knowledge through four-option multiple-choice questions.Mentioned 8 times above

More from See also

• Gemini 3 Pro
• Gemma
• Apple Intelligence
• Llama 3.2
• LLaMA
• Mistral 7B
• Qwen
• GPT-4o
• Small language model
• On-device AI
• Edge AI
• Knowledge Distillation
• Mixture of Experts (MoE)
• TensorFlow Lite (LiteRT)
• Tensor Processing Unit (TPU)
• KV Cache
• Edge computing
• Samsung AI
• Qualcomm
• Sundar Pichai
• Demis Hassabis
• Google DeepMind

Topics:GoogleLarge Language Models

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Pichai, Sundar and Hassabis, Demis, "Introducing Gemini: our largest and most capable AI model", Google, 2023-12-06. blog.google/...google-gemini-ai. Accessed 2026-05-24.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22Gemini Team, Google, "Gemini: A Family of Highly Capable Multimodal Models", arXiv:2312.11805, 2023-12-19. arxiv.org/...2312.11805. Accessed 2026-05-24.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Google, "Pixel 8 Pro the first smartphone with AI built in is now running Gemini Nano, plus more AI updates coming to the Pixel portfolio", Google, 2023-12-06. blog.google/...pixel-feature-drop-december-2023. Accessed 2026-06-24.
4. ^Li, Abner, "Google rolling out Gemini Nano to Pixel 8 Pro with Android AICore", 9to5Google, 2023-12-06. 9to5google.com/...pixel-8-pro-gemini-nano. Accessed 2026-05-24.
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Android Developers, "A New Foundation for AI on Android", Android Developers Blog, 2023-12-06. android-developers.googleblog.com/...ai-on-android. Accessed 2026-05-24.
6. ^TechCrunch, "Pixel 8 Pro becomes the first smartphone powered by Google's new AI model, Gemini", TechCrunch, 2023-12-06. techcrunch.com/...d-by-googles-new-ai-model-gemini. Accessed 2026-05-24.
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Google, "Gemini Nano on Android", Android Developers, 2026. developer.android.com/...gemini-nano. Accessed 2026-05-24.
8. ^1 ^2 ^3 ^4 ^5 ^6Android Developers, "An introduction to privacy and safety for Gemini Nano", Android Developers Blog, 2024-10. android-developers.googleblog.com/...y-gemini-nano. Accessed 2026-05-24.
9. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Android Developers, "Gemini Nano is now available on Android via experimental access", Android Developers Blog, 2024-10-01. android-developers.googleblog.com/...le-on-android. Accessed 2026-05-24.
10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Android Developers, "On-device GenAI APIs as part of ML Kit help you easily build with Gemini Nano", Android Developers Blog, 2025-05. android-developers.googleblog.com/...t-gemini-nano. Accessed 2026-05-24.
11. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Wankhede, Calvin, "Here are all the Gemini Nano features, and the phones that support them", Android Authority, 2024-10-13. androidauthority.com/...o-features-devices-3490062. Accessed 2026-05-24.
12. ^1 ^2 ^3 ^4 ^5Google, "The power of Google AI comes to the new Samsung Galaxy S24 series", Google, 2024-01-17. blog.google/...google-ai-samsung-galaxy-s24. Accessed 2026-05-24.
13. ^1 ^2 ^3Li, Abner, "Google Messages getting on-device Magic Compose with Gemini Nano", 9to5Google, 2024-01-17. 9to5google.com/...new-google-messages-features-beta. Accessed 2026-05-24.
14. ^Google, "Gemini Nano Multimodal Capabilities on Pixel Phones", Google Store, 2024. store.google.com/...gemini-nano-offline. Accessed 2026-05-24.
15. ^1 ^2Schoon, Ben, "Google bringing Gemini Nano to Pixel 8 with next Feature Drop", 9to5Google, 2024-03-28. 9to5google.com/...pixel-8-gemini-nano-feature-drop. Accessed 2026-05-24.
16. ^1 ^2 ^3 ^4 ^5Sharma, Adamya, "Pixel 9 series will debut a more powerful, multimodal Gemini Nano", Android Authority, 2024-05-14. androidauthority.com/...multimodal-pixel-9-3442273. Accessed 2026-05-24.
17. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Li, Abner, "How Pixel Recorder is using Gemini Nano with Multimodality, which is approximately 2x larger than 1.0", 9to5Google, 2024-09-01. 9to5google.com/...corder-gemini-nano-multimodality. Accessed 2026-06-24.
18. ^1 ^2 ^3 ^4Daley, Dean, "Google's Pixel 9a has a smaller, less capable Gemini", MobileSyrup, 2025-03-20. mobilesyrup.com/...e-pixel-9a-runs-gemini-nano-xxs. Accessed 2026-05-24.
19. McHugh-Johnson, Molly, "14 new things you can do with Pixel thanks to AI", Google, 2024-08-13. blog.google/...google-pixel-9-new-ai-features. Accessed 2026-05-24.
20. ^1 ^2 ^3 ^4 ^5 ^6 ^7McHugh-Johnson, Molly, "14 new things you can do with Pixel thanks to AI" (Pixel Studio, Pixel Screenshots, Call Notes and Pixel Weather sections), Google, 2024-08-13. blog.google/...google-pixel-9-new-ai-features. Accessed 2026-05-24.
21. ^1 ^2 ^3Apple, "Introducing Apple Intelligence", Apple Newsroom, 2024-06-10. apple.com/...-intelligence-for-iphone-ipad-and-mac. Accessed 2026-05-24.
22. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Apple Machine Learning Research, "Introducing Apple's On-Device and Server Foundation Models", Apple, 2024-06. machinelearning.apple.com/...ple-foundation-models. Accessed 2026-05-24.
23. ^1 ^2 ^3 ^4 ^5Abdin, Marah et al, "Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone", arXiv:2404.14219, 2024-04. arxiv.org/...2404.14219. Accessed 2026-05-24.
24. ^Google, "Gemini Nano on Android: Building with on-device gen AI", Google I/O 2025. io.google/...technical-session-14. Accessed 2026-05-24.
25. ^1 ^2 ^3 ^4 ^5 ^6McHugh-Johnson, Molly, "5 new things Gemini can do on Pixel", Google, 2025-08-21. blog.google/...gemini-nano-pixel-10-updates. Accessed 2026-05-24.
26. ^1 ^2 ^3 ^4 ^5 ^6 ^7Google, "Pixel 10 introduces new chip, Tensor G5", Google, 2025-08. blog.google/...tensor-g5-pixel-10. Accessed 2026-05-24.
27. ^1 ^2 ^3 ^4Williams, Jaime, "Gemini is coming to your Wear OS smartwatch", Google, 2025-07-09. blog.google/...gemini-wear-os-watches. Accessed 2026-05-24.
28. ^Google Developers, "Blazing fast on-device GenAI with LiteRT-LM", Google Developers Blog, 2026-05-19. developers.googleblog.com/...-genai-with-litert-lm. Accessed 2026-05-24.
29. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20Google, "The Prompt API", Chrome for Developers, 2025-05-20 (updated 2026-10-21). developer.chrome.com/...prompt-api. Accessed 2026-06-24.
30. ^1 ^2 ^3 ^4 ^5 ^6Chrome for Developers, "AI APIs are in stable and origin trials, with new Early Preview Program APIs", Chrome for Developers Blog, 2025-05. developer.chrome.com/...ai-api-updates-io25. Accessed 2026-05-24.
31. ^Westenberg, Mishaal, "Exclusive: Google Pixel 9's Tensor G4 is the smallest upgrade to the series so far", Android Authority, 2024-08. androidauthority.com/...r-g4-small-upgrade-3466398. Accessed 2026-05-24.
32. ^Li, Abner, "Gemini Intelligence has high Android spec requirements, likely won't support Pixel 9 or Galaxy Z Fold 7", 9to5Google, 2026-05-15. 9to5google.com/...igence-android-spec-requirements. Accessed 2026-05-24.
33. ^ML Journey, "Gemini AI Model Parameters and Performance Benchmarks", ML Journey, 2024. mljourney.com/...meters-and-performance-benchmarks. Accessed 2026-05-24.
34. ^1 ^2Google Developers, "On-device GenAI in Chrome, Chromebook Plus, and Pixel Watch with LiteRT-LM", Google Developers Blog, 2025-09-24. developers.googleblog.com/...-watch-with-litert-lm. Accessed 2026-11-05.
35. ^1 ^2InfoQ, "Google Brings Gemini Nano to ML Kit with New On-Device GenAI APIs", InfoQ, 2025-06. infoq.com/...google-mlkit-genai-gemini-nano. Accessed 2026-05-24.
36. ^Google. "Gemini logo (Google Image Library)". Google Image Library publication permission. Original official raster, unchanged.
37. ^1 ^2Android Developers, "Announcing Gemma 4 in the AICore Developer Preview", Android Developers Blog, 2026-04-02. android-developers.googleblog.com/...loper-Preview. Accessed 2026-11-05.
38. ^1 ^2 ^3Li, Abner, "Gemini Intelligence requires Gemini Nano 3, Google lists first Nano 4 phones", 9to5Google, 2026-08-27. 9to5google.com/...gemini-intelligence-nano-4. Accessed 2026-11-05.
39. ^1 ^2 ^3Li, Abner, "Google did more than just switch to TSMC for Tensor G5 - This is what's new", 9to5Google, 2025-08-20. 9to5google.com/...pixel-10-tensor-g5. Accessed 2026-11-05.

Show all 39 references