The global small language models market was valued at USD 2.35 billion in 2025. It is estimated at USD 3.10 billion in 2026 and is projected to reach USD 25.59 billion by 2036, growing at a CAGR of 23.5% over the 2026-2036 forecast period. The report covers model technology and benchmarking, business models, hardware and memory supply, regulation, the value chain, investment and M&A, and competition.
The study covers language models with about 15 billion parameters or fewer that are built to run efficiently on phones, PCs, vehicles, edge devices, and on-premises servers, or cheaply in the cloud, along with the platforms and services used to adapt and deploy them. In scope are proprietary and open-weight families such as Microsoft Phi, Google Gemma, Alibaba Qwen, IBM Granite, and TII Falcon, on-device models from device makers such as Apple, adaptation and inference tools, and related services. Market revenue comes from model licenses and APIs, paid support for open-weight models, platform fees, OEM deals, and services. Frontier-scale models and general-purpose AI processors fall outside the scope.
According to a June 2025 paper by NVIDIA Research and the Georgia Institute of Technology, serving a 7-billion-parameter model costs 10 to 30 times less in latency, energy, and compute than serving a model with 70 to 175 billion parameters, and specialized small models could take over 40% to 70% of large-model queries in the popular open-source agents the authors studied. Device platforms are also opening up: Apple reported that its Foundation Models framework, launched at WWDC 2025, gives developers offline access to an on-device model of about 3 billion parameters, compressed to 2-bit precision and available in 15 languages, without inference fees. The UAE's Technology Innovation Institute reported more than 55 million downloads of its Falcon models.
Memory shortages limit growth, because AI data centers absorb so much high-bandwidth memory that standard DRAM and NAND are in short supply; in December 2025, Micron said it would quit consumer memory, Samsung raised some memory chip prices by up to 60% the same month, and both Samsung and SK hynix reported 2026 orders above their production capacity, which pushes device makers to cap memory in mid-range phones and PCs. Small models still trail frontier models in open-ended conversation and complex reasoning, so they often work alongside rather than instead of large models. Many capable models are free to download, which shifts revenue toward tooling and services, and the EU AI Act has applied documentation, copyright, and transparency duties to general-purpose model providers since 2 August 2025. Other challenges include fragmented chips and runtimes, scarce labeled data and MLOps skills, weak low-resource language coverage, and checking the origin of open-weight models.
Hybrid systems that send most tasks to small models and call large models only when needed are the largest opportunity, benefiting suppliers of routing, orchestration, and evaluation platforms and of specialist models for common agent tasks. Edge and industrial uses such as robots, vehicles, and field equipment need compact language and vision-language models that work without a reliable connection. Regulated sectors such as banking, healthcare, and government favor models they can host and audit themselves. Key trends include hybrid Transformer and state-space designs such as TII's Falcon-H1, aggressive quantization and flash-based weight loading to save memory, and distillation of reasoning from frontier models into smaller ones.
By model size, the report covers models below 1B, 1B-4B, 4B-10B, and 10B-15B parameters. The 4B-10B class is expected to hold the largest share in 2026, as it gives enterprise and agent workloads enough capability at a manageable cost; it includes the 7-billion-parameter class highlighted by NVIDIA Research and TII's Falcon 3. The 1B-4B class is expected to grow at the highest CAGR, driven by deployment on smartphones, PCs, and wearables, where models such as Apple's roughly 3-billion-parameter on-device model must fit within tight memory and power budgets.
The offering analysis spans models (proprietary licensing and APIs and commercial support for open-weight models), platforms & tools (fine-tuning and distillation, quantization and compression, evaluation and guardrails, and inference runtimes), and services (customization and integration and managed services). Platforms & tools are expected to account for the largest share in 2026, because the wide availability of free open-weight models moves spending away from model access and toward the software needed to adapt, compress, test, and run models in production. Services are expected to grow at the highest CAGR, as enterprises without in-house data pipelines and fine-tuning skills turn to partners to build specialized models and run them under managed contracts.
By deployment, the market is divided into on-device (smartphones & tablets, PCs & laptops, wearables & XR devices, and vehicles), edge & on-premises, and cloud. Cloud is expected to hold the largest share in 2026, as most enterprise agent and API workloads are still served from cloud infrastructure, where small models lower the cost per call. On-device deployment is expected to grow at the highest CAGR, supported by operating system frameworks from Apple and on-device models from Google that let millions of apps use local AI without per-request fees.
By modality, the report compares text models with multimodal models, including vision-language and speech models. Text models are expected to hold the larger share in 2026, since most current workloads, such as customer service, document processing, coding, and search, are text based. Multimodal models are expected to grow faster, as robots, vehicles, wearables, and smartphones need models that can interpret images and speech locally.
By application, the report covers agentic AI & tool calling, customer service & virtual assistants, code generation, document processing & summarization, search & RAG, translation & localization, and IoT & robotics. Customer service & virtual assistants are expected to account for the largest share in 2026, reflecting high volumes of routine queries in which fast, low-cost responses matter more than broad world knowledge. Agentic AI & tool calling is expected to grow at the highest CAGR, as the cost gap identified by NVIDIA Research encourages enterprises to move repetitive agent steps to fine-tuned small models.
The end-use industry analysis covers IT & telecom, BFSI, healthcare & life sciences, retail & e-commerce, manufacturing & automotive, government & defense, consumer electronics, and education. IT & telecom is expected to hold the largest share in 2026, as technology firms and network operators are early adopters of AI agents, coding assistants, and customer support automation. Healthcare & life sciences is expected to grow at the highest CAGR, driven by demand for private deployment that keeps sensitive patient data within the organization's control. BFSI and government & defense also favor self-hosted models.
North America is expected to be the largest regional market in 2026, as Microsoft, Google, Meta, Apple, NVIDIA, IBM, and other major developers of small models are based in the U.S. and American enterprises are early adopters of agentic AI, although device makers such as Dell have raised prices because of memory costs. Canada adds Cohere and strong AI research institutes. Europe holds a significant share, shaped by the EU AI Act, France-based Mistral AI, and demand for on-premises and sovereign deployment in banking, healthcare, and public administration, and regulators there scrutinize model origins, as when Italy's privacy watchdog blocked DeepSeek's app in early 2025.
Asia-Pacific is expected to grow at the highest CAGR during the forecast period, led by China, which supplies many of the most widely used open-weight small models, notably Alibaba's Qwen family (also the base for widely used distilled models), and hosts large device makers rolling out on-device AI. South Korea's memory makers shape device costs worldwide, Japan is investing in Japanese-language models, and the IndiaAI Mission is backing sovereign models built for Indian languages. In the Middle East & Africa, the UAE's Falcon models and Saudi Arabia's national AI programs support strong growth, while in Latin America, Chile's National Center for Artificial Intelligence leads the Latam-GPT initiative and Brazil and Mexico have large developer communities.
The market is highly competitive and fragmented, as large technology companies release small models next to their frontier models, device makers build their own on-device models, and independent developers, national research institutes, and a wide range of platform and tooling suppliers compete for enterprise use. Because many leading models have open weights, rivalry turns on how much capability each parameter delivers, hardware efficiency, license terms, supported languages, ecosystem fit, and surrounding tools and services. Leading players release model families across sizes, tune them for specific chips, distill reasoning from larger models, and partner with device makers and enterprises. Recent developments include Apple's Foundation Models framework in 2025 and its third-generation models in June 2026, and TII's May 2025 launch of Falcon Arabic and the Falcon-H1 family.
Key companies profiled in the report include Microsoft Corporation (U.S.), Google LLC (U.S.), Meta Platforms, Inc. (U.S.), Apple Inc. (U.S.), Alibaba Cloud (China), Mistral AI (France), NVIDIA Corporation (U.S.), IBM Corporation (U.S.), Hugging Face, Inc. (U.S.), Technology Innovation Institute (UAE), DeepSeek (China), Cohere Inc. (Canada), Liquid AI, Inc. (U.S.), Salesforce, Inc. (U.S.), Qualcomm Incorporated (U.S.), ModelBest Inc. (China), Sarvam AI (India), and Arcee AI (U.S.).
The study covers language models with about 15 billion parameters or fewer that are built to run efficiently on phones, PCs, vehicles, edge devices, and on-premises servers, or cheaply in the cloud, along with the platforms and services used to adapt and deploy them. In scope are proprietary and open-weight families such as Microsoft Phi, Google Gemma, Alibaba Qwen, IBM Granite, and TII Falcon, on-device models from device makers such as Apple, adaptation and inference tools, and related services. Market revenue comes from model licenses and APIs, paid support for open-weight models, platform fees, OEM deals, and services. Frontier-scale models and general-purpose AI processors fall outside the scope.
According to a June 2025 paper by NVIDIA Research and the Georgia Institute of Technology, serving a 7-billion-parameter model costs 10 to 30 times less in latency, energy, and compute than serving a model with 70 to 175 billion parameters, and specialized small models could take over 40% to 70% of large-model queries in the popular open-source agents the authors studied. Device platforms are also opening up: Apple reported that its Foundation Models framework, launched at WWDC 2025, gives developers offline access to an on-device model of about 3 billion parameters, compressed to 2-bit precision and available in 15 languages, without inference fees. The UAE's Technology Innovation Institute reported more than 55 million downloads of its Falcon models.
Market Dynamics
Agentic AI is the strongest driver, as enterprise agents make many repetitive calls for classification, extraction, and tool use, and methods such as LoRA let a small model be specialized in GPU-hours, so routing these calls to small models cuts costs sharply. A second driver is the arrival of small models inside operating systems; Apple released a third generation of its foundation models at WWDC26 in June 2026, and Google offers on-device models for smartphones, letting app makers add generative features without paying for cloud inference. Sovereign AI adds a third driver, as governments and national institutes build models in local languages that they can run on their own infrastructure. TII's Falcon Arabic, released in May 2025, matched models up to ten times larger on Arabic benchmarks.Memory shortages limit growth, because AI data centers absorb so much high-bandwidth memory that standard DRAM and NAND are in short supply; in December 2025, Micron said it would quit consumer memory, Samsung raised some memory chip prices by up to 60% the same month, and both Samsung and SK hynix reported 2026 orders above their production capacity, which pushes device makers to cap memory in mid-range phones and PCs. Small models still trail frontier models in open-ended conversation and complex reasoning, so they often work alongside rather than instead of large models. Many capable models are free to download, which shifts revenue toward tooling and services, and the EU AI Act has applied documentation, copyright, and transparency duties to general-purpose model providers since 2 August 2025. Other challenges include fragmented chips and runtimes, scarce labeled data and MLOps skills, weak low-resource language coverage, and checking the origin of open-weight models.
Hybrid systems that send most tasks to small models and call large models only when needed are the largest opportunity, benefiting suppliers of routing, orchestration, and evaluation platforms and of specialist models for common agent tasks. Edge and industrial uses such as robots, vehicles, and field equipment need compact language and vision-language models that work without a reliable connection. Regulated sectors such as banking, healthcare, and government favor models they can host and audit themselves. Key trends include hybrid Transformer and state-space designs such as TII's Falcon-H1, aggressive quantization and flash-based weight loading to save memory, and distillation of reasoning from frontier models into smaller ones.
Segment Analysis
The report analyzes the market by model size, offering, deployment, modality, application, end-use industry, and geography, showing where small models create the most value.By model size, the report covers models below 1B, 1B-4B, 4B-10B, and 10B-15B parameters. The 4B-10B class is expected to hold the largest share in 2026, as it gives enterprise and agent workloads enough capability at a manageable cost; it includes the 7-billion-parameter class highlighted by NVIDIA Research and TII's Falcon 3. The 1B-4B class is expected to grow at the highest CAGR, driven by deployment on smartphones, PCs, and wearables, where models such as Apple's roughly 3-billion-parameter on-device model must fit within tight memory and power budgets.
The offering analysis spans models (proprietary licensing and APIs and commercial support for open-weight models), platforms & tools (fine-tuning and distillation, quantization and compression, evaluation and guardrails, and inference runtimes), and services (customization and integration and managed services). Platforms & tools are expected to account for the largest share in 2026, because the wide availability of free open-weight models moves spending away from model access and toward the software needed to adapt, compress, test, and run models in production. Services are expected to grow at the highest CAGR, as enterprises without in-house data pipelines and fine-tuning skills turn to partners to build specialized models and run them under managed contracts.
By deployment, the market is divided into on-device (smartphones & tablets, PCs & laptops, wearables & XR devices, and vehicles), edge & on-premises, and cloud. Cloud is expected to hold the largest share in 2026, as most enterprise agent and API workloads are still served from cloud infrastructure, where small models lower the cost per call. On-device deployment is expected to grow at the highest CAGR, supported by operating system frameworks from Apple and on-device models from Google that let millions of apps use local AI without per-request fees.
By modality, the report compares text models with multimodal models, including vision-language and speech models. Text models are expected to hold the larger share in 2026, since most current workloads, such as customer service, document processing, coding, and search, are text based. Multimodal models are expected to grow faster, as robots, vehicles, wearables, and smartphones need models that can interpret images and speech locally.
By application, the report covers agentic AI & tool calling, customer service & virtual assistants, code generation, document processing & summarization, search & RAG, translation & localization, and IoT & robotics. Customer service & virtual assistants are expected to account for the largest share in 2026, reflecting high volumes of routine queries in which fast, low-cost responses matter more than broad world knowledge. Agentic AI & tool calling is expected to grow at the highest CAGR, as the cost gap identified by NVIDIA Research encourages enterprises to move repetitive agent steps to fine-tuned small models.
The end-use industry analysis covers IT & telecom, BFSI, healthcare & life sciences, retail & e-commerce, manufacturing & automotive, government & defense, consumer electronics, and education. IT & telecom is expected to hold the largest share in 2026, as technology firms and network operators are early adopters of AI agents, coding assistants, and customer support automation. Healthcare & life sciences is expected to grow at the highest CAGR, driven by demand for private deployment that keeps sensitive patient data within the organization's control. BFSI and government & defense also favor self-hosted models.
Regional Analysis
The report covers North America, Europe, Asia-Pacific, Latin America, and the Middle East & Africa, with analysis of 20 countries, the Nordic countries, Australia & New Zealand, Southeast Asia, and four rest-of-region groups. It considers model development, device and cloud ecosystems, enterprise adoption, regulation, and sovereign AI programs.North America is expected to be the largest regional market in 2026, as Microsoft, Google, Meta, Apple, NVIDIA, IBM, and other major developers of small models are based in the U.S. and American enterprises are early adopters of agentic AI, although device makers such as Dell have raised prices because of memory costs. Canada adds Cohere and strong AI research institutes. Europe holds a significant share, shaped by the EU AI Act, France-based Mistral AI, and demand for on-premises and sovereign deployment in banking, healthcare, and public administration, and regulators there scrutinize model origins, as when Italy's privacy watchdog blocked DeepSeek's app in early 2025.
Asia-Pacific is expected to grow at the highest CAGR during the forecast period, led by China, which supplies many of the most widely used open-weight small models, notably Alibaba's Qwen family (also the base for widely used distilled models), and hosts large device makers rolling out on-device AI. South Korea's memory makers shape device costs worldwide, Japan is investing in Japanese-language models, and the IndiaAI Mission is backing sovereign models built for Indian languages. In the Middle East & Africa, the UAE's Falcon models and Saudi Arabia's national AI programs support strong growth, while in Latin America, Chile's National Center for Artificial Intelligence leads the Latam-GPT initiative and Brazil and Mexico have large developer communities.
Competitive Landscape
The report provides a detailed review of the competitive landscape, covering key growth strategies, competitive benchmarking, a competitive dashboard that groups companies into industry leaders, market differentiators, vanguards, and emerging companies, and a 2025 market share/rank analysis, together with company profiles that include business overview, financial overview, product portfolio, strategic developments, and SWOT analysis.The market is highly competitive and fragmented, as large technology companies release small models next to their frontier models, device makers build their own on-device models, and independent developers, national research institutes, and a wide range of platform and tooling suppliers compete for enterprise use. Because many leading models have open weights, rivalry turns on how much capability each parameter delivers, hardware efficiency, license terms, supported languages, ecosystem fit, and surrounding tools and services. Leading players release model families across sizes, tune them for specific chips, distill reasoning from larger models, and partner with device makers and enterprises. Recent developments include Apple's Foundation Models framework in 2025 and its third-generation models in June 2026, and TII's May 2025 launch of Falcon Arabic and the Falcon-H1 family.
Key companies profiled in the report include Microsoft Corporation (U.S.), Google LLC (U.S.), Meta Platforms, Inc. (U.S.), Apple Inc. (U.S.), Alibaba Cloud (China), Mistral AI (France), NVIDIA Corporation (U.S.), IBM Corporation (U.S.), Hugging Face, Inc. (U.S.), Technology Innovation Institute (UAE), DeepSeek (China), Cohere Inc. (Canada), Liquid AI, Inc. (U.S.), Salesforce, Inc. (U.S.), Qualcomm Incorporated (U.S.), ModelBest Inc. (China), Sarvam AI (India), and Arcee AI (U.S.).
How This Report Helps
- Provides reliable market size estimates and long-term forecasts for the global small language models market through 2036.
- Explains how agentic AI economics, on-device AI platforms, and sovereign AI programs are creating demand for small models.
- Assesses the effect of memory shortages, capability limits, free open-weight models, and the EU AI Act on market growth.
- Identifies high-growth opportunities across model size, offering, deployment, modality, application, end-use industry, and region.
- Tracks technology trends in hybrid architectures, quantization, distillation, fine-tuning, on-device runtimes, and model routing.
- Benchmarks leading model developers, platform providers, and chipmakers on model portfolios, ecosystem reach, and partnerships.
- Supports product strategy, platform investment, partnership, and market entry decisions for AI developers, device makers, enterprises, and investors.
Key Questions Answered
- How large is the global small language models market today, and how will it grow through 2036?
- Why are small models better suited than frontier models to many agentic AI workloads?
- What are the main drivers, restraints, opportunities, challenges, and trends shaping the market?
- How are device platforms such as Apple's Foundation Models framework changing on-device AI adoption?
- Which model sizes, offerings, deployment modes, modalities, applications, and industries will grow the fastest?
- How will memory costs and hardware fragmentation affect on-device small models?
- How can vendors earn revenue when many capable small models are free to download?
- Which regions and countries offer the most attractive opportunities, and how are sovereign AI programs shaping demand?
- Who are the leading companies in the market, and what strategies are they using to compete?
Table of Contents
1. Introduction
2. Research Methodology
4. Market Overview
5. Small Language Models Market, by Model Size
6. Small Language Models Market, by Offering
7. Small Language Models Market, by Deployment
8. Small Language Models Market, by Modality
9. Small Language Models Market, by Application
10. Small Language Models Market, by End-Use Industry
11. Small Language Models Market, by Geography
12. Competitive Landscape
13. Company Profiles
14. Appendix
LIST OF TABLES
LIST OF FIGURES
Companies Mentioned
- Microsoft Corporation
- Google LLC
- Meta Platforms, Inc.
- Apple Inc.
- Alibaba Cloud
- Mistral AI
- NVIDIA Corporation
- IBM Corporation
- Hugging Face, Inc.
- Technology Innovation Institute
- DeepSeek
- Cohere Inc.
- Liquid AI, Inc.
- Salesforce, Inc.
- Qualcomm Incorporated
- ModelBest Inc.
- Sarvam AI
- Arcee AI
