GenAI-capable smartphones are on track to reach 45% of global shipments in 2026, up from 36% a year earlier (Counterpoint Research, GenAI Smartphone Share Forecast 2026), and nearly all of them depend on a small language model running locally on a neural processing unit. The models themselves have shrunk to fit: Apple’s on-device foundation model packs about 3 billion parameters into 2 bits per weight (Apple Machine Learning Research, Foundation Models 2025 Updates), and Google’s Gemma 3n E2B runs in as little as 2GB of memory (Google Developers Blog, Gemma 3n Developer Guide 2025). Silicon is moving just as fast, with Qualcomm claiming up to 220 tokens per second on its newest Hexagon NPU. We aggregated data from Counterpoint Research, Gartner, IDC, MLCommons, Apple, Google, Qualcomm, MediaTek, Meta, Microsoft, Deloitte and Samsung to map where on-device SLMs stand in 2026. For the wider hardware picture, see our edge AI statistics roundup.
TL;DR
- 45% of smartphones shipped in 2026 will be GenAI-capable, up from 36% in 2025 (Counterpoint Research, 2026)
- 559 million GenAI smartphones forecast for 2026, after 369.3 million in 2025 (Gartner, GenAI Smartphone Forecast 2025)
- $393.3 billion in end-user spending on GenAI smartphones forecast for 2026, up 32% (Gartner, 2025)
- About 3 billion parameters at 2 bits per weight power Apple’s on-device model (Apple Machine Learning Research, 2025)
- 2GB and 3GB memory footprints for Gemma 3n E2B and E4B (Google Developers Blog, 2025)
- 56% smaller and 41% less memory for quantized Llama 3.2 1B and 3B models (Meta AI, 2024)
- Up to 220 tokens per second on the Snapdragon 8 Elite Gen 5 Hexagon NPU, 37% faster than the prior generation (Qualcomm, 2025)
- 2.6x faster and 2x more efficient Gemini Nano on Tensor G5 versus Tensor G4 (Google, 2025)
- 3 Llama model sizes (1B, 3B, 8B) in the new MLPerf Mobile v6.0 LLM benchmark (MLCommons, 2026)
- Roughly fourfold rise in mobile memory prices by Q2 2026 (Counterpoint Research, 2026)
- 13.9% drop in 2026 smartphone shipments to 1.08 billion units forecast because of the memory squeeze (Counterpoint Research, 2026)
- More than 70% of Galaxy S25 users actively engage with built-in AI features (Samsung, 2025)
1. GenAI Smartphone Shipments: The Installed Base for On-Device SLMs
Every GenAI smartphone is, by definition, a host for a small language model: Gartner defines the category as devices with a built-in neural engine or NPU capable of running SLMs. That makes shipment forecasts the cleanest proxy for how many devices can run a local model. Counterpoint now expects 45% of all smartphones shipped in 2026 to be GenAI-capable, and 52% in 2027 (Counterpoint Research, GenAI Smartphone Share to Rise to 45% in 2026).
The story behind the curve is the mid-range. Counterpoint says GenAI is now standard above $400 wholesale. The constraint has moved from “can the chip run a model” to “can the bill of materials absorb the memory”, which Section 6 quantifies. If you want the broader handset context, our smartphone market statistics page covers total volumes by vendor.
| Metric | Value | Source |
|---|---|---|
| GenAI-capable share of global smartphone shipments, 2026 | 45% (up from 36% in 2025) | Counterpoint Research, June 2026 |
| GenAI-capable share forecast, 2027 | 52% | Counterpoint Research, June 2026 |
| GenAI smartphone shipments, 2025 forecast | More than 400 million, about one in three phones | Counterpoint Research, March 2025 |
| GenAI-capable share in 2024 | One in five smartphones | Counterpoint Research, March 2025 |
| GenAI smartphone units, 2024 / 2025 / 2026 | 260.4M / 369.3M / 559M | Gartner, GenAI Smartphone Forecast, Sept 2025 |
| GenAI smartphone shipments, 2025 | Over 370 million (30% share), up 73.1% | IDC Worldwide Smartphone Forecast, 2025 |
| GenAI smartphone shipments, 2028 | 912 million units, 78.4% CAGR | IDC, Generative AI Smartphone Forecast, July 2024 |
Context note: the three firms use different definitions and dates. IDC’s July 2024 release (IDC press release) projected GenAI phones at 70% of the market by 2028, while its 2025 forecast pushed the “over 70%” mark to 2029. Most recent IDC long-range data: July 2024.
2. Spending and Market Size: Big Hardware Dollars, Small Model Revenue
Two very different numbers describe this market. The hardware that runs SLMs is a mainstream consumer category, while the market for the models themselves is still small and badly defined. Gartner forecasts $393.3 billion in end-user spending on GenAI smartphones in 2026, up 32% from $298.2 billion in 2025 (Gartner, GenAI Smartphone End-User Spending press release 2025).
On the software side, analyst estimates for the small language model market disagree by an order of magnitude, which tells you more about definitions than about demand. MarketsandMarkets sizes 2025 at $0.93 billion while Research and Markets puts it near $9.16 billion. The common thread: every firm flags edge and on-device deployment as the fastest-growing segment.
| Metric | Value | Source |
|---|---|---|
| End-user spending on GenAI smartphones, 2024 / 2025 / 2026 | $244.7B / $298.2B / $393.3B | Gartner, Sept 2025 |
| GenAI smartphones as share of all AI end-user spending, 2025 | 20% | Gartner, Sept 2025 |
| Basic (under $350) GenAI smartphones with NPUs, 2025 | 41% | Gartner, Sept 2025 |
| Premium smartphones with GenAI capability by 2029 | 100% | Gartner, Sept 2025 |
| Small language model market, 2025 to 2032 | $0.93B to $5.45B, 28.7% CAGR | MarketsandMarkets, April 2025 |
| Small language model market, 2025 to 2030 | $9.16B to $22.45B, 19.6% CAGR | Research and Markets, SLM Market Report 2026 |
| Small language model market, 2026 to 2030 | $11.1B to $20.7B, 15.1% CAGR | Grand View Research, SLM Market Report |
MarketsandMarkets’ release (GlobeNewswire, April 2025) names edge deployment as the fastest-growing segment, driven by latency and privacy demands on smartphones and IoT devices.
3. Model Size and Compression: 3 Billion Parameters at 2 Bits
The practical ceiling for a phone-resident model is set by RAM, not ambition. Shipping models cluster between 1B and 4B parameters, and the real engineering is in how few bits each weight gets. Apple compresses its roughly 3-billion-parameter on-device model to 2 bits per weight using quantization-aware training, with 4-bit embeddings and an 8-bit KV cache (Apple Machine Learning Research, Updates to Apple’s On-Device and Server Foundation Language Models 2025).
That is a sharp step down from 2024, when Apple’s model averaged 3.7 bits per weight. Google took a different route with Gemma 3n, keeping more raw parameters but parking per-layer embeddings outside accelerator memory so only about 2B or 4B “effective” parameters sit in fast memory (Google Developers Blog, Introducing Gemma 3n 2025). Compression costs little accuracy at these levels: Apple reports a 1.5% MMLU improvement and about a 4.6% MGSM regression after compression.
| Metric | Value | Source |
|---|---|---|
| Apple on-device model size | About 3 billion parameters | Apple Machine Learning Research, 2025 |
| Apple on-device weight precision | 2 bits per weight (QAT); embeddings 4-bit; KV cache 8-bit | Apple Machine Learning Research, 2025 |
| Apple KV cache memory reduction from block sharing | 37.5% | Apple Intelligence Foundation Language Models Tech Report, 2025 |
| Apple on-device model precision in 2024 | Mixed 2-bit and 4-bit, averaging 3.7 bits per weight | Apple Machine Learning Research, 2024 |
| Gemma 3n E2B / E4B raw parameters | 5B / 8B | Google Developers Blog, 2025 |
| Gemma 3n E2B / E4B memory footprint | As little as 2GB / 3GB | Google Developers Blog, 2025 |
| Quantized Llama 3.2 1B and 3B size and memory reduction vs BF16 | 56% smaller, 41% less memory | Meta AI, Quantized Llama Models, Oct 2024 |
| Phi Silica size and quantization | 3.3B parameters, 4-bit weights, about 60% memory reduction | Microsoft Windows Blog, Dec 2024 |
Context note: Meta’s quantized Llama release (Meta AI blog) and Microsoft’s Phi Silica post (Windows Experience Blog) date from late 2024. Most recent available data for both: 2024. A laptop-based academic study adds a useful rule of thumb: around 3.5 effective bits per weight is where a heavily quantized larger model starts beating a smaller, higher-precision one (arXiv 2505.15030, A Systematic Evaluation of On-Device LLMs).
4. NPU Throughput: Tokens per Second on Flagship Silicon
Throughput claims are vendor-reported and measured on different models, so treat them as ceilings rather than comparable scores. Even so, the direction is unmistakable. Qualcomm says the Snapdragon 8 Elite Gen 5 Hexagon NPU is 37% faster than its predecessor and sustains up to 220 tokens per second, with INT2 and FP8 support (Qualcomm, Snapdragon 8 Elite Gen 5 launch, Sept 2025; figures corroborated by Futurum Group).
The more telling numbers are prefill versus decode. Reading a prompt is cheap on an NPU; generating each new token is memory-bound. Gemma 3n E2B on a Dimensity 9500 NPU reads more than 1,600 tokens per second but writes 28, which is why every vendor is fighting for bandwidth and lower-precision formats. That gap matters for real-time voice and assistant features, where a fast first response hides a slow tail; our voice assistant statistics track how that shows up in usage.
| Metric | Value | Source |
|---|---|---|
| Snapdragon 8 Elite Gen 5 NPU uplift | 37% faster than prior generation | Qualcomm, Sept 2025 |
| Snapdragon 8 Elite Gen 5 sustained generation | Up to 220 tokens per second | Qualcomm, Sept 2025 |
| Snapdragon 8 Elite Gen 5 overall SoC power savings | 16% | Futurum Group analysis of Qualcomm data, Oct 2025 |
| Dimensity 9500 3B-parameter LLM output | 2x faster (100%), with 56% lower peak power | MediaTek Dimensity 9500 launch, Sept 2025 |
| Gemma 3n E2B on Dimensity 9500 NPU (4K context) | 1,600+ tokens/s prefill, 28 tokens/s decode | Google LiteRT NeuroPilot results, Dec 2025 |
| LiteRT NPU throughput vs CPU / GPU on MediaTek | Up to 12x CPU, 10x GPU | Google LiteRT NeuroPilot results, Dec 2025 |
| Gemini Nano on Tensor G5 vs Tensor G4 | 2.6x faster, 2x more efficient; TPU up to 60% more powerful | Google, Pixel 10 Tensor G5 announcement, Aug 2025 |
| Apple on-device model (iPhone 15 Pro) | 0.6 ms per prompt token to first token; 30 tokens/s | Apple Machine Learning Research, 2024 |
MediaTek’s figures come from its launch release (MediaTek via PR Newswire); Google’s Tensor G5 claims come from the Pixel 10 Tensor G5 post. Apple’s tokens-per-second figure is the most recent it has published: 2024.
5. Benchmarks and Quality: MLPerf Mobile Enters the LLM Era
Until 2026, there was no neutral yardstick for phone LLMs; every number in Section 4 is a vendor’s own. MLPerf Mobile v6.0, released by MLCommons on June 15, 2026, adds LLM benchmarks built on Llama 3.2 1B, Llama 3.2 3B and Llama 3.1 8B, scored on TinyMMLU and IFEval (MLCommons, MLPerf Mobile v6.0).
The inclusion of an 8B model is the headline: MLCommons notes v6.0 supports NPU-accelerated execution of Llama 3.1 8B on Snapdragon 8 Elite Gen 5, a size that was a data-center benchmark a year earlier. On quality, small models are now close to their bigger peers on knowledge tests. Apple’s 3B on-device model scores 67.85% on MMLU, slightly ahead of Qwen-2.5-3B, and Phi Silica loses less than one point on PIQA after 4-bit quantization. For the open-weight families behind many of these benchmarks, see our open-source LLM statistics.
| Metric | Value | Source |
|---|---|---|
| MLPerf Mobile v6.0 LLM models | Llama 3.2 1B, Llama 3.2 3B, Llama 3.1 8B | MLCommons, June 2026 |
| SoCs with new or updated v6.0 support | Snapdragon 8 Elite Gen 5, Dimensity 9500, Exynos 2600 | MLCommons, June 2026 |
| MLPerf Mobile reference quality, Llama 3.1 8B | TinyMMLU about 61%, IFEval about 67% (FP16) | MLCommons, MLPerf Mobile benchmark page |
| MLPerf Mobile reference quality, Llama 3.2 3B | TinyMMLU about 59%, IFEval about 65.5% (FP16) | MLCommons, MLPerf Mobile benchmark page |
| MLPerf Inference 5.1 small-LLM interactive limits (Llama 3.1 8B) | First token within 0.5 s, 30 ms per output token | MLCommons, Sept 2025 |
| Apple on-device model MMLU | 67.85% vs Qwen-2.5-3B 66.37% and Gemma-3-4B 62.81% | Apple Foundation Models Tech Report, 2025 |
| Phi Silica accuracy after 4-bit quantization | PIQA 80.47% to 79.76%; ARC Challenge 63.48% to 60.49% | Microsoft Windows Blog, Dec 2024 |
| Gemma 3n E4B LMArena score | Over 1300 | Google Developers Blog, 2025 |
Context note: the MLPerf Mobile benchmark page (MLCommons) uses 90th-percentile latency as its primary metric for single-stream tasks. Public device-level LLM results from v6.0 had not been published in the release announcement we reviewed. The Apple scores are from the 2025 tech report on arXiv.
6. Memory and Cost Constraints: The DRAM Squeeze
The biggest threat to on-device SLMs in 2026 is not model quality; it is the price of the RAM they live in. Counterpoint reports mobile memory prices rose roughly fourfold by Q2 2026 as AI data centers absorbed most DRAM and NAND output (Counterpoint Research, Mobile Memory Cost Hikes, Sept 2026).
Watch how the forecast deteriorated in six months: a 2.1% shipment decline in December 2025, 12.4% in March 2026, and 13.9% by June. A 3B model at 2 bits needs well under a gigabyte for weights, but the multi-gigabyte RAM headroom that GenAI features expect is exactly what budget phones are now cutting. That explains why Counterpoint ties the next wave of GenAI phones to both memory supply and more efficient models.
| Metric | Value | Source |
|---|---|---|
| Mobile memory price change by Q2 2026 | Roughly fourfold | Counterpoint Research, Sept 2026 |
| Mobile LPDDR4/5 price in Q2 2026 vs Q3 2025 | Nearly 3x | Counterpoint Research, via EE Times Asia, March 2026 |
| BoM cost increase from DRAM surge, low / mid / high-end | About 25% / 15% / 10% | Counterpoint Research, Dec 2025 |
| Wholesale ASP revision, 2026 | Up 6.9% year over year | Counterpoint Research, Dec 2025 |
| 2026 smartphone shipment forecast, Dec 2025 / Mar 2026 / Jun 2026 | -2.1% / -12.4% / -13.9% | Counterpoint Research |
| 2026 global smartphone shipments | 1.08 billion units, lowest on record | Counterpoint Research, June 2026 |
| Sub-$200 segment shipments, 2026 | Decline of more than 20% | Counterpoint Research, via EE Times Asia, March 2026 |
Sources: Counterpoint, 2026 Smartphone Shipment Forecasts Revised Down and EE Times Asia on the March 2026 revision. Gartner adds a hardware floor: it expects NPUs above 40 TOPS to be standard in premium phones by 2027.
7. Adoption, Usage and Trust: Developers and Consumers
Distribution is no longer the bottleneck. Apple opened its on-device model to third-party apps through the Foundation Models framework, with free inference that works offline and needs as few as three lines of Swift (Apple Newsroom, WWDC25). Samsung says more than 70% of Galaxy S25 users actively engage with built-in AI features, and it targeted Galaxy AI on over 400 million devices by the end of 2025, double its 200 million goal for 2024 (Samsung Newsroom, Galaxy Tech Forum 2025).
Trust is the counterweight, and it is the strongest argument for keeping inference local. Deloitte found 70% of US consumers worry about data privacy and security, and only 27% have high confidence their data is protected. On-device SLMs answer that concern by design, which is also why local processing matters for voice tools such as a real-time voice changer for Discord, where audio never needs to leave the machine.
| Metric | Value | Source |
|---|---|---|
| Galaxy S25 users actively engaging with built-in AI | More than 70% | Samsung, July 2025 |
| Galaxy AI device target, end of 2025 | Over 400 million (vs 200 million in 2024) | Samsung, July 2025 |
| Users saying AI features are part of daily routines | 47% | Samsung / Symmetry Research survey, 2025 |
| Users who use voice commands as often as they type | 45% | Samsung / Symmetry Research survey, 2025 |
| Consumers planning to buy a GenAI smartphone within a year | 59% | Counterpoint Research consumer survey, Dec 2024 |
| US consumers who use or experiment with GenAI | 53% | Deloitte Connected Consumer Survey 2025 (n=3,524) |
| US consumers worried about data privacy and security | 70% | Deloitte Connected Consumer Survey 2025 |
| Consumers with high confidence their data is protected | 27% | Deloitte Connected Consumer Survey 2025 |
Survey details: Deloitte 2025 Connected Consumer press release and Counterpoint upgrade-intent survey. Counterpoint’s survey is from December 2024, the most recent available upgrade-intent data from the firm.
Summary: On-Device Small Language Models by the Numbers
| Metric | Value | Source |
|---|---|---|
| GenAI-capable share of smartphone shipments, 2026 | 45% | Counterpoint Research, 2026 |
| GenAI-capable share, 2027 | 52% | Counterpoint Research, 2026 |
| GenAI smartphone units, 2026 | 559 million | Gartner, 2025 |
| GenAI smartphone end-user spending, 2026 | $393.3 billion (+32%) | Gartner, 2025 |
| GenAI smartphone shipments, 2028 | 912 million | IDC, 2024 |
| Small language model market, 2032 | $5.45 billion (28.7% CAGR) | MarketsandMarkets, 2025 |
| Apple on-device model | About 3B parameters at 2 bits per weight | Apple Machine Learning Research, 2025 |
| Apple KV cache memory reduction | 37.5% | Apple Tech Report, 2025 |
| Gemma 3n E2B memory footprint | As little as 2GB | Google Developers Blog, 2025 |
| Quantized Llama 3.2 size / memory reduction | 56% / 41% | Meta AI, 2024 |
| Snapdragon 8 Elite Gen 5 NPU | Up to 220 tokens/s, 37% faster | Qualcomm, 2025 |
| Dimensity 9500 3B LLM output | 2x faster, 56% lower peak power | MediaTek, 2025 |
| Gemma 3n E2B on Dimensity 9500 NPU | 1,600+ tokens/s prefill, 28 tokens/s decode | Google LiteRT, 2025 |
| Gemini Nano on Tensor G5 | 2.6x faster, 2x more efficient | Google, 2025 |
| MLPerf Mobile v6.0 LLM sizes | 1B, 3B, 8B | MLCommons, 2026 |
| Mobile memory prices by Q2 2026 | Roughly fourfold | Counterpoint Research, 2026 |
| 2026 smartphone shipments | 1.08 billion (-13.9%) | Counterpoint Research, 2026 |
| Galaxy S25 users engaging with AI | More than 70% | Samsung, 2025 |
| US consumers worried about privacy | 70% | Deloitte, 2025 |
Methodology and Sources
- Counterpoint Research, GenAI Smartphone Share to Rise to 45% of Global Shipments in 2026 (June 2026), with segment detail reported by Electronics For You
- Counterpoint Research, GenAI Smartphone Shipments to Exceed 400 Million in 2025 (March 2025)
- Counterpoint Research, 2026 Smartphone Shipment Forecasts Revised Down as Memory Shortage Drives BoM Costs Up (December 2025)
- Counterpoint Research, Mobile Memory Cost Hikes (September 2026)
- Counterpoint Research, consumer survey on GenAI smartphone upgrades (December 2024)
- EE Times Asia, report on Counterpoint’s March 2026 shipment revision
- Gartner, Worldwide GenAI Smartphone End-User Spending press release (September 2025)
- IDC, Worldwide Generative AI Smartphone Shipments Forecast (July 2024) and IDC Worldwide Smartphone Forecast (2025)
- MarketsandMarkets, Small Language Model Market release (April 2025); Research and Markets, Small Language Model Market Report 2026; Grand View Research, Small Language Model Market Report
- Apple Machine Learning Research, Updates to Apple’s On-Device and Server Foundation Language Models (2025), Apple Intelligence Foundation Language Models Tech Report 2025 (arXiv), Introducing Apple’s On-Device and Server Foundation Models (2024) and Apple Newsroom, WWDC25
- Google Developers Blog, Introducing Gemma 3n (June 2025) and Google, Pixel 10 Tensor G5 (August 2025)
- Google LiteRT NeuroPilot results on Dimensity 9500 (December 2025), as reported by MarkTechPost
- Qualcomm, Snapdragon 8 Elite Gen 5 launch (September 2025) and Futurum Group analysis
- MediaTek, Dimensity 9500 launch release (September 2025)
- Meta AI, Introducing Quantized Llama Models (October 2024)
- Microsoft, Phi Silica, small but mighty on-device SLM (December 2024)
- MLCommons, MLPerf Mobile v6.0 (June 2026), MLPerf Mobile benchmark page and MLPerf Inference 5.1 small LLM benchmark (September 2025)
- arXiv 2505.15030, A Systematic Evaluation of On-Device LLMs: Quantization, Performance, and Resources
- Samsung Newsroom, Galaxy Tech Forum at Galaxy Unpacked 2025 (July 2025), survey figures as reported by FoneArena
- Deloitte, 2025 Connected Consumer Survey press release (September 2025)
- Data watch: 2025 GenAI smartphone volumes differ by firm: IDC said over 370 million (30% share), Gartner 369.3 million, and Counterpoint more than 400 million in its March 2025 forecast, later restating the 2025 share at 36%. For 2026, Counterpoint’s 45% of 1.08 billion units implies about 486 million GenAI phones (0.45 x 1.08 billion), well below Gartner’s 559 million, which was published in September 2025, before the memory shortage forced shipment cuts; expect Gartner to revise. Small language model market sizing ranges from $0.93 billion (MarketsandMarkets) to about $9 to $11 billion (Research and Markets, Grand View Research) for 2025-2026 because of differing definitions. All NPU throughput figures are vendor-reported on different models, precisions and context lengths, so they are not directly comparable; MLPerf Mobile v6.0 device-level LLM results are the data to watch. Apple’s 30 tokens/s figure, Meta’s quantized Llama results, Microsoft’s Phi Silica figures and IDC’s 2028 forecast are 2024 data, the most recent each source has published. Apple’s ML research page lists 15 supported languages while the arXiv tech report describes 16 locales.
Last updated: October 3, 2026. We update this roundup quarterly, and the next refresh is expected when Counterpoint publishes its Q3 2026 smartphone and memory price trackers and MLCommons releases the first public MLPerf Mobile v6.0 LLM results.