Gemini Nano: On-Device Gemini Model Guide 2026 (v3)
Gemini Nano v3 (2026) runs 1.8B-3.25B params fully on-device via Android AICore with zero API cost, sub-second latency, and a 4,096-token context window.
Gemini Nano is Google DeepMind's on-device model line, launched December 6, 2023 and now at v3 (2026) alongside Android 17, running 1.8B (Nano-1) to 3.25B (Nano-2) parameters locally via AICore with a roughly 4,096-token context window. It requires zero API cost and 12GB+ RAM on qualifying flagship devices like the Pixel 10 series.
Gemini Nano is Google DeepMind's on-device model, first released December 6, 2023 and now on v3 (2026), running 1.8B to 3.25B parameters entirely on Android hardware via AICore. It costs $0 per inference and delivers sub-second latency, but its context window is limited to about 4,096 tokens.
Provider: Google DeepMind · Family: Gemini Nano
Context window: 4,096 tokens · Max output: 1,024
Input modalities: text, image, audio · Output: text
About Gemini Nano
Gemini Nano is Google DeepMind's on-device model line, the smallest tier in the Gemini family, first released December 6, 2023 alongside Gemini Pro and Ultra, and shipped initially on the Pixel 8 Pro. Unlike every other Gemini variant, Nano does not run on Google's servers: it executes entirely on the user's phone through Android's AICore system service, using 4-bit quantization to fit two model sizes, Nano-1 at 1.8B parameters for lower-memory devices and Nano-2 at 3.25B parameters for flagship hardware. By Google I/O 2026, Google had shipped Gemini Nano v3 alongside Android 17, calling it the most capable on-device model Google has shipped inside a smartphone, and a research team separately published work retrofitting Multi-Token Prediction onto frozen Nano v3 weights to speed up inference without retraining. Because Nano is a fundamentally different product category from cloud Gemini models, its benchmark story is different too: it is not chasing SWE-bench or GPQA Diamond scores against GPT-5 or Claude Opus, it is competing against other on-device models like Apple's on-device foundation model and Microsoft's Phi-3 family on the tradeoff between quality-per-parameter and latency. Independent testing shows Gemini 1.5 Pro leads pure MMLU among Google's own lineup but with 1-3 second cloud round-trip latency, while Apple's on-device model responds in under 1 second; Gemini Nano's own strength is that it matches or beats that sub-second latency profile on supported Pixel and Samsung hardware while running fully offline. Community testers report Nano still fails classic small-model traps like counting letters in "strawberry," a known artifact of aggressive parameter reduction shared across the entire sub-4B on-device model class, not unique to Google. Context handling is deliberately constrained for the hardware target: Gemini Nano's effective context is roughly 4,096 tokens, with Google's own guidance recommending prompts stay under 1,024 tokens for reliable results, dramatically smaller than the 1M-token windows on cloud Gemini 3 models. This is a hardware and latency tradeoff, not an oversight: a larger context window on a 1.8B-3.25B parameter model would blow the memory and thermal budget of a phone. Modalities depend on the exposed API surface rather than raw model capability: Google's ML Kit GenAI APIs (built on AICore) expose Nano for text-only and some multimodal prompts, image description generation, on-device speech-to-text transcription, summarization, proofreading, and message rewriting in different tones. Full raw multimodal input/output specifics beyond these packaged APIs are not published in detail by Google. Pricing is the model's standout feature: there is no per-token cost at all. Inference runs on the user's own processor (Tensor TPU, Snapdragon NPU, or MediaTek APU depending on device), so there are zero server costs, zero network calls, and zero API bills for developers building on it, a structurally different economics model from every cloud LLM. The tradeoff is hardware gatekeeping: Gemini Intelligence (the 2026 feature bundle built on Nano v3) requires at least 12GB of RAM and a current-generation flagship chipset, confirmed on Pixel 10 series, Galaxy S26 series, OnePlus 15 series, OPPO Find X9/X8, and select 2026 Xiaomi, realme, Honor, and Motorola flagships, meaning older devices like Pixel 9 or Galaxy S25 cannot run the newest tier. Deployment is exclusively on-device through the AICore system service and exposed to third-party developers via ML Kit's GenAI APIs and an AICore Developer Preview program that allows bypassing production quota limits for testing. There is no cloud API, no Vertex AI listing, and no direct model download for Gemini Nano; it ships baked into supported Android builds. In production, apps can hit a "BUSY" error under frequent testing (a deliberate device-protection throttle) and the very first inference after a cold start can take up to about a minute while the model loads into memory. Safety and data handling are the other headline differentiator: because inference never leaves the device, there are no server-side data retention concerns, no network transmission of prompts, and no cloud provider access to user content, positioning Nano as Google's answer to on-device privacy for sensitive personal-assistant tasks (message drafting, call summarization, on-device transcription) where users are reluctant to send data to any cloud API, including Google's own. Gemini Nano is the right choice for developers building privacy-sensitive Android features that need to work offline with zero marginal cost per inference: message assistance, on-device summarization, real-time transcription, and lightweight image description. It is the wrong choice for anything needing large context, complex multi-step reasoning, code generation at a professional standard, or cross-platform reach beyond a narrow set of qualifying Android flagships; those workloads belong on cloud Gemini 3 Pro/Flash or a competing frontier API.
Pricing
No API pricing exists because there is no server-side inference. Cost to developers is zero; the only cost is the end user's device hardware, which must meet Google's RAM and chipset requirements to run Nano v3 / Gemini Intelligence.
Key Features
- Fully On-Device Inference: Runs entirely on the phone's NPU/TPU via Android AICore, so prompts and outputs never leave the device and there is no network dependency.
- Zero Marginal Cost: No per-token or per-request pricing exists since there is no server-side inference; cost to developers is $0 regardless of call volume.
- Two Hardware-Matched Sizes: Ships as Nano-1 (1.8B params) for lower-memory devices and Nano-2 (3.25B params) for flagship hardware, both 4-bit quantized to fit mobile memory budgets.
- ML Kit GenAI Task APIs: Packaged APIs for summarization, proofreading, tone rewriting, image description, and on-device speech-to-text, built directly on top of AICore.
- AICore Developer Preview: Lets developers bypass production quota throttling during testing to avoid the BUSY error that protects devices in normal use.
Pros
- Zero per-inference cost at any volume, since inference runs on the user's own device hardware rather than a billed cloud endpoint.
- Sub-second latency on supported flagship hardware, with no network round-trip since processing never leaves the device.
- No server-side data retention or transmission, a real privacy advantage for sensitive personal-assistant features.
Cons
- Effective context window of roughly 4,096 tokens (recommended under 1,024) is far below cloud Gemini 3 models, ruling out long-document tasks.
- Gemini Intelligence on Nano v3 requires 12GB+ RAM and a current flagship chipset, locking out devices as recent as Pixel 9 and Galaxy S25.
- No cloud API or cross-platform equivalent exists; it is exclusively an Android on-device feature with no iOS or web access path.
Frequently Asked Questions
What is Gemini Nano and who built it?
Gemini Nano is Google DeepMind's smallest Gemini model tier, built specifically to run entirely on-device rather than on Google's cloud servers. It first launched December 6, 2023 alongside Gemini Pro and Ultra, shipping initially on the Pixel 8 Pro, and has since progressed to Gemini Nano v3, launched alongside Android 17 at Google I/O 2026 and described by Google as the most capable on-device model it has shipped inside a smartphone. It comes in two hardware-matched sizes: Nano-1 at 1.8B parameters for lower-memory devices and Nano-2 at 3.25B parameters for flagship hardware, both using 4-bit quantization to fit mobile memory budgets. Unlike cloud Gemini models competing on SWE-bench or GPQA Diamond, Nano's competitive axis is on-device latency and zero-cost inference against rivals like Apple's on-device foundation model and Microsoft's Phi-3 family. It runs through Android's AICore system service and has no equivalent cloud API listing.
How much does Gemini Nano cost per 1M tokens?
Gemini Nano has no per-token or per-request API pricing at all, because inference runs entirely on the end user's device hardware rather than on a billed cloud endpoint. There are no network calls, no server-side compute charges, and no API costs for developers regardless of call volume, whether that's 100 calls or 100 million calls across an app's user base. The only cost is indirect: the end user's device must meet Google's hardware bar (12GB+ RAM and a current flagship chipset for the newest Gemini Intelligence features on Nano v3), which is a device-purchase cost, not a per-inference API cost. This zero-marginal-cost structure is fundamentally different from every cloud LLM's per-token billing model and is Nano's single biggest differentiator for high-volume consumer features. There is no batch API, no cached-input discount, and no enterprise pricing tier, because there is no billed API surface to discount.
What is Gemini Nano's context window and max output?
Gemini Nano's effective context window is roughly 4,096 tokens, and Google's own guidance recommends keeping prompts under 1,024 tokens for reliable results. This is dramatically smaller than the 1 million token context windows on cloud Gemini 3 models, and it's a deliberate hardware tradeoff rather than an oversight: a larger context window on a 1.8B-3.25B parameter model would exceed the memory and thermal budget of a phone. There is no separate extended-context tier for Nano, unlike cloud Gemini variants. Long documents, multi-file inputs, or extended conversation history are not use cases Nano is built for; those workloads should route to cloud Gemini 3 Pro or Flash instead. Google's packaged ML Kit GenAI APIs (summarization, proofreading, rewriting) are tuned to work within this constrained window rather than exposing the full raw context budget to developers.
How does Gemini Nano compare to Apple Intelligence and Phi-3 on-device models?
Independent testing shows the on-device model class trades raw benchmark quality for latency and footprint rather than chasing frontier scores: Apple's on-device foundation model responds in under 1 second, and Gemini Nano targets a comparable sub-second latency profile on supported Pixel and Samsung hardware. Microsoft's Phi-3.5-MoE reportedly scores around 78% on MMLU, competitive with cloud models like Gemini 1.5 Flash and GPT-4o mini, though that is a larger MoE model rather than a strict phone-resident model like Nano. Community testers report that all small on-device models, Gemini Nano included, still fail classic small-model traps such as counting letters in "strawberry," a known artifact of aggressive parameter reduction across the entire sub-4B model class rather than a Gemini-specific weakness. Nano's practical edge over both rivals is deep Android OS integration via AICore and zero API cost, while Apple's model has equivalent OS-level integration on iOS and Phi-3 has the advantage of being available as open weights for cross-platform deployment.
Is Gemini Nano open source or proprietary?
Gemini Nano is proprietary and is not distributed as downloadable open weights; it ships bundled inside Android OS builds and Google Play services rather than being published to Hugging Face or under an open license. Access for third-party developers is exclusively through Android's AICore system service and Google's ML Kit GenAI task APIs (summarization, proofreading, rewriting, image description, speech-to-text), with an AICore Developer Preview program available for testing with relaxed quota limits. There is no Vertex AI listing, no direct API key access, and no self-hostable weight file, unlike Google's open-weight Gemma line, which is a separate, unrelated product. Commercial use in an Android app is permitted under standard Android developer and Google Play terms, since developers are calling a packaged system service rather than licensing a model file directly.
What modalities does Gemini Nano support?
Through Google's ML Kit GenAI APIs, Gemini Nano exposes text-based prompting, some multimodal image input for image description generation, and on-device audio input for speech-to-text transcription; there is no audio output or video input/output capability published for this tier. Function calling, structured JSON output, and tool use are not exposed at the ML Kit API level the way they are on cloud Gemini 3 models, since the packaged task APIs (summarize, proofread, rewrite, describe image, transcribe) are the primary integration surface rather than a raw completion endpoint. Google has not published detailed raw multimodal specifications beyond what these packaged APIs expose, so developers needing a specific modality combination should verify support against the current ML Kit GenAI documentation rather than assuming full Gemini 3 parity.
Does Gemini Nano train on user data?
No training-on-user-data concern applies to Gemini Nano in the way it does for cloud APIs, because inference runs entirely on the user's device and prompts or outputs never transmit to Google's servers at all. There is no data retention policy to configure because there is no server-side storage of inputs or outputs in the first place; the model file is downloaded to the device once and runs locally from then on. This makes Nano Google's strongest privacy story across its entire model lineup, since even zero-retention cloud API tiers still involve a network transmission that Nano avoids entirely. Google states this on-device processing model supports GDPR compliance by design, since no personal data leaves the device for this specific inference path. There is no published SOC 2, ISO 27001, or HIPAA-eligible certification for Nano specifically since those attestations apply to cloud infrastructure that Nano does not use.
Who is Gemini Nano best for and who should avoid it?
It is best for Android developers building privacy-sensitive personal-assistant features, offline-capable summarization or transcription tools, and any high-volume consumer feature where zero marginal inference cost matters more than frontier-level reasoning quality. It is the wrong choice for long-document analysis or large multi-file context, since its roughly 4,096-token window is a small fraction of cloud Gemini 3's 1M-token capacity; those workloads should use Gemini 3 Pro or Flash instead. It is also the wrong choice for complex multi-step reasoning or professional-grade code generation, where a 1.8B-3.25B parameter model simply lacks the capacity of frontier cloud models. Teams needing cross-platform reach should avoid depending on it entirely, since Nano ships exclusively through Android's AICore with no iOS, web, or general cloud API equivalent, and even within Android, the newest Gemini Intelligence tier on Nano v3 requires 12GB+ RAM and a current flagship chipset, excluding devices as recent as Pixel 9 and Galaxy S25.