Gemini Nano fits Android developers who need offline, privacy-sensitive features like message drafting, call summarization, or on-device transcription that never touch a server. It replaces a cloud API call for those tasks, but Gemini Intelligence on Nano v3 needs 12GB+ RAM and a current flagship chipset, so it only runs on recent hardware like the Pixel 10 series.
Google DeepMind built Gemini Nano as its smallest Gemini tier, an on-device model now on version 3 (2026) alongside Android 17. It is Google's on-device answer to Apple's on-device foundation model, running entirely inside Android's AICore service with sub-second latency and full offline operation unlike Google's larger cloud Gemini models.
Where it sits
- $0/M$ per 1M tokensBlended price (3:1)Lower is better#1 / 64peer median $1.70/Mvendor price, checked by HokAI
- --tokens/sOutput speedHigher is better-- / 39peer median 90 tok/scited: Artificial Analysis
- --% solvedSWE-bench VerifiedHigher is better-- / 28peer median 78.3%per source, see benchmark scores
- --% correctGPQA DiamondHigher is better-- / 44peer median 88.3%per source, see benchmark scores
Cheaper than 97% of the 64 GA models with a published price, and one of 65 whose vendor states it does not train on customer data.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: Google DeepMind · Family: Gemini Nano
More about Google DeepMind on HokAI
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 never touches Google's servers: it runs entirely on the phone through Android's AICore system service. By Google I/O 2026, Google had shipped Gemini Nano v3 alongside Android 17, calling it the most capable on-device model it 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.
Nano's competitive set isn't cloud frontier models chasing SWE-bench or GPQA Diamond scores, it's other on-device models like Apple's on-device foundation model and Microsoft's Phi-3 family, where the tradeoff is quality-per-parameter against latency rather than raw capability. Independent testing shows Apple's model responding in under 1 second; Nano targets that same sub-second profile on qualifying Pixel and Samsung hardware while running fully offline, though testers report it still misses precise character-level tasks the way most sub-4B on-device models do.
Because Nano is a phone-resident model, both its small context window and its RAM floor are hardware-driven rather than product choices: a wider context or lower device-RAM requirement on a model this size would blow a phone's memory and thermal budget. That same tradeoff explains why there is no cloud API, Vertex AI listing, or downloadable weights: Nano ships baked into qualifying Android builds and reaches third-party developers only through ML Kit's GenAI APIs and the AICore Developer Preview program.
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.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0 | $0 | $0 |
| Support reply | $0 | $0 | $0 |
| One coding agent run | $0 | $0 | $0 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
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 does Gemini Nano actually cost?
Gemini Nano has no per-token or per-request API pricing because inference runs entirely on the end user's device instead of a billed cloud endpoint, so the cost to a developer is $0 no matter how many calls the app makes. The only real cost is indirect: the user's phone has to clear Google's hardware bar for Nano v3's Gemini Intelligence features, which is a device-purchase cost, not an API fee. There's no volume discount, cached-input rate, or enterprise contract to negotiate, since there is no billed API surface to discount at all.
Can you use Gemini Nano without paying?
Yes, and there is no separate free plan because there is no paid plan at all: Nano runs at no cost regardless of call volume, since every request stays on the device instead of hitting a metered cloud endpoint. The catch isn't a price gate, it's a hardware gate: Gemini Intelligence, the 2026 feature bundle on Nano v3, requires a current flagship-tier phone that meets Google's RAM and chipset bar, so an older or budget device simply cannot access it no matter what a developer or user is willing to pay.
What are the best alternatives to Gemini Nano?
The closest comparisons are Google's own Gemma 4 12B, an open-weight model with a 256K-token context window that developers can self-host outside of Android entirely, and Microsoft's Phi-4, a 14B open-source model that is MIT licensed for broad reuse. Apple's on-device foundation model competes on the same on-device, sub-second-latency pitch but is locked to iOS the way Nano is locked to Android. None of the three matches Nano's specific mix of no per-token billing and native AICore integration, available only to developers already building for Android.
How does Gemini Nano compare to Phi-4 in 2026?
Phi-4 is a 14B open-source model, MIT licensed, that scores 84.8% on MMLU and runs on a single consumer GPU or through Azure AI Foundry's pay-per-token API; Gemini Nano is a smaller proprietary model that only runs inside Android's AICore service. The practical split is platform and cost: Phi-4 needs a GPU or an Azure API bill but runs on any operating system, while Nano needs a qualifying Android phone and then runs at no cost per call at all. Nano's advantage only shows up inside that narrow Android-only, offline-first use case; outside it, Phi-4's open license and cross-platform reach win out.
What does it take to start using Gemini Nano?
Developers integrate through Google's ML Kit GenAI APIs (summarization, proofreading, tone rewriting, image description, on-device speech-to-text) rather than a raw prompt endpoint, so the fastest path is calling one of those packaged task APIs instead of hand-rolling a prompt. During testing, enable the AICore Developer Preview's quota bypass to avoid the BUSY throttle error, but remember that setting applies to every app on the device, so turn it off again before handing the phone to a real user. Expect the very first call after a cold start to take up to about a minute while the model loads into memory, so warm it in the background rather than triggering it on first user interaction.
Top Alternatives
- Gemma 4 12B: Pick Gemma 4 12B if you need an open-weight model you can self-host with a 256K-token context window; pick Gemini Nano if you want zero-cost, zero-setup inference already built into Android.
- Phi-4: Pick Phi-4 if you want a model that runs on a consumer GPU or through Azure's paid API; pick Gemini Nano if your target is an Android phone and inference has to cost nothing.