List of the Top Small Language Models for Startups in 2026 - Page 3

Amazon Nova signifies a groundbreaking advancement in foundation models (FMs), delivering sophisticated intelligence and exceptional price-performance ratios, exclusively accessible through Amazon Bedrock. The series features Amazon Nova Micro, Amazon Nova Lite, and Amazon Nova Pro, each tailored to process text, image, or video inputs and generate text outputs, addressing varying demands for capability, precision, speed, and operational expenses. Amazon Nova Micro is a model centered on text, excelling in delivering quick responses at an incredibly low price point. On the other hand, Amazon Nova Lite is a cost-effective multimodal model celebrated for its rapid handling of image, video, and text inputs. Lastly, Amazon Nova Pro distinguishes itself as a powerful multimodal model that provides the best combination of accuracy, speed, and affordability for a wide range of applications, making it particularly suitable for tasks like video summarization, answering queries, and solving mathematical problems, among others. These innovative models empower users to choose the most suitable option for their unique needs while experiencing unparalleled performance levels in their respective tasks. This flexibility ensures that whether for simple text analysis or complex multimodal interactions, there is an Amazon Nova model tailored to meet every user's specific requirements.

Phi-4 is an innovative small language model (SLM) with 14 billion parameters, demonstrating remarkable proficiency in complex reasoning tasks, especially in the realm of mathematics, in addition to standard language processing capabilities. Being the latest member of the Phi series of small language models, Phi-4 exemplifies the strides we can make as we push the horizons of SLM technology. Currently, it is available on Azure AI Foundry under a Microsoft Research License Agreement (MSRLA) and will soon be launched on Hugging Face. With significant enhancements in methodologies, including the use of high-quality synthetic datasets and meticulous curation of organic data, Phi-4 outperforms both similar and larger models in mathematical reasoning challenges. This model not only showcases the continuous development of language models but also underscores the important relationship between the size of a model and the quality of its outputs. As we forge ahead in innovation, Phi-4 serves as a powerful example of our dedication to advancing the capabilities of small language models, revealing both the opportunities and challenges that lie ahead in this field. Moreover, the potential applications of Phi-4 could significantly impact various domains requiring sophisticated reasoning and language comprehension.

Qwen2.5-VL-32B is a sophisticated AI model designed for multimodal applications, excelling in reasoning tasks that involve both text and imagery. This version builds upon the advancements made in the earlier Qwen2.5-VL series, producing responses that not only exhibit superior quality but also mirror human-like formatting more closely. The model excels in mathematical reasoning, in-depth image interpretation, and complex multi-step reasoning challenges, effectively addressing benchmarks such as MathVista and MMMU. Its capabilities have been substantiated through performance evaluations against rival models, often outperforming even the larger Qwen2-VL-72B in particular tasks. Additionally, with enhanced abilities in image analysis and visual logic deduction, Qwen2.5-VL-32B provides detailed and accurate assessments of visual content, allowing it to formulate insightful responses based on intricate visual inputs. This model has undergone rigorous optimization for both text and visual tasks, making it exceptionally adaptable to situations that require advanced reasoning and comprehension across diverse media types, thereby broadening its potential use cases significantly. As a result, the applications of Qwen2.5-VL-32B are not only diverse but also increasingly relevant in today's data-driven landscape.

Amazon Nova Micro is a high-performance, text-only AI model that provides low-latency responses, making it ideal for applications needing real-time processing. With impressive capabilities in language understanding, translation, and reasoning, Nova Micro can generate over 200 tokens per second while maintaining high performance. This model supports fine-tuning on text inputs and is highly efficient, making it perfect for cost-conscious businesses looking to deploy AI for fast, interactive tasks such as code completion, brainstorming, and solving mathematical problems.

Amazon Nova Lite is an efficient multimodal AI model built for speed and cost-effectiveness, handling image, video, and text inputs seamlessly. Ideal for high-volume applications, Nova Lite provides fast responses and excellent accuracy, making it well-suited for tasks like interactive customer support, content generation, and media processing. The model supports fine-tuning on diverse input types and offers a powerful solution for businesses that prioritize both performance and budget.

Phi-4-reasoning is a sophisticated transformer model that boasts 14 billion parameters, crafted specifically to address complex reasoning tasks such as mathematics, programming, algorithm design, and strategic decision-making. It achieves this through an extensive supervised fine-tuning process, utilizing curated "teachable" prompts and reasoning examples generated via o3-mini, which allows it to produce detailed reasoning sequences while optimizing computational efficiency during inference. By employing outcome-driven reinforcement learning techniques, Phi-4-reasoning is adept at generating longer reasoning pathways. Its performance is remarkable, exceeding that of much larger open-weight models like DeepSeek-R1-Distill-Llama-70B, and it closely rivals the more comprehensive DeepSeek-R1 model across a range of reasoning tasks. Engineered for environments with constrained computing resources or high latency, this model is refined with synthetic data sourced from DeepSeek-R1, ensuring it provides accurate and methodical solutions to problems. The efficiency with which this model processes intricate tasks makes it an indispensable asset in various computational applications, further enhancing its significance in the field. Its innovative design reflects an ongoing commitment to pushing the boundaries of artificial intelligence capabilities.

Phi-4-reasoning-plus is an enhanced reasoning model that boasts 14 billion parameters, significantly improving upon the capabilities of the original Phi-4-reasoning. Utilizing reinforcement learning, it achieves greater inference efficiency by processing 1.5 times the number of tokens that its predecessor could manage, leading to enhanced accuracy in its outputs. Impressively, this model surpasses both OpenAI's o1-mini and DeepSeek-R1 on various benchmarks, tackling complex challenges in mathematical reasoning and high-level scientific questions. In a remarkable feat, it even outshines the much larger DeepSeek-R1, which contains 671 billion parameters, in the esteemed AIME 2025 assessment, a key qualifier for the USA Math Olympiad. Additionally, Phi-4-reasoning-plus is readily available on platforms such as Azure AI Foundry and HuggingFace, streamlining access for developers and researchers eager to utilize its advanced features. Its cutting-edge design not only showcases its capabilities but also establishes it as a formidable player in the competitive landscape of reasoning models. This positions Phi-4-reasoning-plus as a preferred choice for users seeking high-performance reasoning solutions.

Phi-4-mini-reasoning is an advanced transformer-based language model that boasts 3.8 billion parameters, tailored specifically for superior performance in mathematical reasoning and systematic problem-solving, especially in scenarios with limited computational resources and low latency. The model's optimization is achieved through fine-tuning with synthetic data generated by the DeepSeek-R1 model, which effectively balances performance and intricate reasoning skills. Having been trained on a diverse set of over one million math problems that vary from middle school level to Ph.D. complexity, Phi-4-mini-reasoning outperforms its foundational model by generating extensive sentences across numerous evaluations and surpasses larger models like OpenThinker-7B, Llama-3.2-3B-instruct, and DeepSeek-R1 in various tasks. Additionally, it features a 128K-token context window and supports function calling, which ensures smooth integration with different external tools and APIs. This model can also be quantized using the Microsoft Olive or Apple MLX Framework, making it deployable on a wide range of edge devices such as IoT devices, laptops, and smartphones. Furthermore, its design not only enhances accessibility for users but also opens up new avenues for innovative applications in the realm of mathematics, potentially revolutionizing how such problems are approached and solved.

Xgen-small is a streamlined language model developed by Salesforce AI Research, specifically designed for enterprise applications, providing effective long-context processing at a reasonable price. It integrates focused data selection, scalable pre-training, extension of context length, instruction-based fine-tuning, and reinforcement learning to meet the sophisticated and high-demand inference requirements of modern enterprises. Unlike traditional large models, Xgen-small stands out in its ability to handle extensive contexts, enabling it to adeptly gather insights from a range of sources, including internal documents, programming code, academic papers, and live data streams. With configurations of 4B and 9B parameters, it achieves a delicate equilibrium between cost-effectiveness, data privacy, and thorough understanding of long contexts, making it a dependable and sustainable choice for extensive Enterprise AI applications. This pioneering method not only boosts operational productivity but also equips organizations with the tools to harness AI effectively in their strategic goals, thus fostering innovation and growth in various sectors. As businesses continue to evolve, solutions like Xgen-small will play a crucial role in shaping the future of AI integration.

OpenAI o4-mini-high offers the performance of a larger AI model in a smaller, more cost-efficient package. With enhanced capabilities in fields like visual perception, coding, and complex problem-solving, o4-mini-high is built for those who require high-throughput, low-latency AI assistance. It's perfect for industries where fast and precise reasoning is critical, such as fintech, healthcare, and scientific research.

On June 23, 2025, Microsoft introduced Mu, a cutting-edge language model boasting 330 million parameters and designed to significantly improve the agent experience in Windows environments by seamlessly converting natural language questions into functional calls for Settings, with all operations executed on-device via NPUs at an impressive speed exceeding 100 tokens per second while maintaining high accuracy. Utilizing Phi Silica optimizations, Mu's encoder-decoder architecture employs a fixed-length latent representation that notably minimizes computational requirements and memory consumption, achieving a 47 percent decrease in first-token latency and delivering a decoding speed that is 4.7 times faster on Qualcomm Hexagon NPUs in comparison to traditional decoder-only models. Furthermore, the model is enhanced by hardware-aware tuning methodologies, which incorporate a strategic 2/3–1/3 division of encoder and decoder parameters, shared weights for both input and output embeddings, Dual LayerNorm, rotary positional embeddings, and grouped-query attention, facilitating rapid inference rates that surpass 200 tokens per second on devices like the Surface Laptop 7, along with response times for settings-related queries that are under 500 ms. This impressive blend of features and optimizations establishes Mu as a revolutionary development in the realm of on-device language processing capabilities, setting new standards for speed and efficiency. As a result, users can expect a more intuitive and responsive experience when interacting with their Windows settings through natural language.

The Phi-4-mini-flash-reasoning model, boasting 3.8 billion parameters, is a key part of Microsoft's Phi series, tailored for environments with limited processing capabilities such as edge and mobile platforms. Its state-of-the-art SambaY hybrid decoder architecture combines Gated Memory Units (GMUs) with Mamba state-space and sliding-window attention layers, resulting in performance improvements that are up to ten times faster and decreasing latency by two to three times compared to previous iterations, while still excelling in complex reasoning tasks. Designed to support a context length of 64K tokens and fine-tuned on high-quality synthetic datasets, this model is particularly effective for long-context retrieval and real-time inference, making it efficient enough to run on a single GPU. Accessible via platforms like Azure AI Foundry, NVIDIA API Catalog, and Hugging Face, Phi-4-mini-flash-reasoning presents developers with the tools to build applications that are both rapid and highly scalable, capable of performing intensive logical processing. This extensive availability encourages a diverse group of developers to utilize its advanced features, paving the way for creative and innovative application development in various fields.

The Nemotron 3 Nano distinguishes itself as the smallest model in NVIDIA's Nemotron 3 series, tailored specifically for agentic AI applications that necessitate strong reasoning and conversational capabilities while ensuring economical inference costs. This innovative hybrid Mamba-Transformer Mixture-of-Experts model is equipped with 3.2 billion active parameters and expands to 3.6 billion when accounting for embeddings, culminating in an impressive total of 31.6 billion parameters. NVIDIA claims that this model achieves superior accuracy compared to its predecessor, the Nemotron 2 Nano, while also operating with less than half of the parameters during each forward pass, thereby boosting efficiency without sacrificing performance. Additionally, it reportedly outperforms both GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507 across a range of commonly used benchmarks. With an input capacity of 8K and an output limit of 16K utilizing a single H200, the model realizes an inference throughput that is 3.3 times higher than that of Qwen3-30B-A3B and 2.2 times that of GPT-OSS-20B. Furthermore, the Nemotron 3 Nano can manage context lengths of up to 1 million tokens, reinforcing its dominance over GPT-OSS-20B and Qwen3-30B-A3B-Instruct-2507. This extraordinary amalgamation of capabilities not only enhances its precision and efficiency but also positions the Nemotron 3 Nano as a premier option for cutting-edge AI endeavors that require top-tier performance. As the demand for advanced AI solutions grows, the relevance of such models will likely continue to expand.

Aion-1.0-Instruct is a recently launched compact language model incorporated into Microsoft Edge as part of a developer preview, which focuses on early testing and collecting user feedback. This innovative model is tailored to improve Edge's on-device Prompt and Writing Assistance APIs, offering web developers a faster, smaller, and more efficient AI-driven solution for browser features. Previously, Microsoft had employed Phi-4-mini for these APIs; however, its high hardware demands limited accessibility across various devices. In contrast, Aion-1.0-Instruct expands compatibility to a significantly wider range of devices, including those with less capable GPUs and even those that operate solely on CPU inference without a GPU, all while preserving excellent performance in various web applications. Developers can access this model through the Edge Canary and Dev channels, allowing them to evaluate its performance in real-world web settings, examine API interoperability, and provide feedback before final modifications. By enabling developers to effortlessly add AI capabilities to their websites and browser extensions, Aion-1.0-Instruct aims to enrich user experiences significantly. Moreover, its introduction could potentially revolutionize web development, making AI features more accessible and user-friendly for a larger audience. As the landscape of web technologies continues to evolve, the implications of this model will likely extend far beyond initial expectations.

SubQ 1.1 Small is a long-context enterprise AI model developed by Subquadratic to address the limitations of traditional models that struggle with large artifacts. It is built for tasks where the full context matters, including analyzing entire codebases, reviewing lengthy contracts, comparing financial filings, and reasoning across document collections. The model uses Subquadratic Sparse Attention, which replaces dense attention with a learned sparse approach that scales more efficiently as context length grows. This allows SubQ 1.1 Small to process extremely large context windows while sharply reducing attention compute requirements. In benchmark testing, the model achieved near-perfect needle-in-a-haystack retrieval at 1M, 2M, 6M, and 12M tokens. It also scored 99.12% on the RULER 128K benchmark, demonstrating strength on tasks involving multi-hop reasoning, variable tracing, aggregation, and long-context understanding. Beyond retrieval, SubQ 1.1 Small maintains competitive performance in general knowledge, coding, and enterprise agent benchmarks such as GPQA Diamond, LiveCodeBench, and AutomationBench Finance. Its efficiency is a major advantage, requiring 64.5x less compute than dense attention and running 56x faster than FlashAttention-2 at 1M tokens on a single attention layer. The model was trained through staged context extension and continued pretraining on long-form artifacts such as books, documents, and repository-scale code. SubQ 1.1 Small is suited for financial analysis, legal work, software engineering, due diligence, long-horizon coding tasks, and enterprise workflows that depend on relationships spread across large bodies of information. It gives organizations a way to reason over complete artifacts more directly instead of relying only on retrieval pipelines, chunking strategies, and agentic scaffolding.

Bonsai 27B emerges as the newest flagship in the Bonsai series, representing the first-ever 27B-class model crafted for mobile device functionality. Leveraging the foundation of Qwen3.6 27B, this model significantly enhances local device capabilities with sophisticated multi-step reasoning, structured tool interactions, vision tasks, and agentic loops that ensure coherence across numerous operations. The Bonsai 27B is offered in two unique versions, with the Ternary Bonsai 27B utilizing ternary weights alongside FP16 group-wise scaling to achieve an effective weight of 1.71 bits, while maintaining a 5.9 GB footprint ideal for high-performance laptops. Alternatively, the 1-bit Bonsai 27B adopts binary weights with the same group-wise scaling approach, resulting in an effective weight of 1.125 bits and a reduced size of 3.9 GB, which aligns well with the memory limitations of devices such as the iPhone 17 Pro. Both variants operate smoothly throughout the entire language network, encompassing embeddings, attention mechanisms, MLPs, and the language model head, without the need for higher-precision solutions. Additionally, they include a compact 4-bit vision tower, which empowers on-device workflows to accurately analyze screenshots, documents, and camera inputs, thereby improving user interaction and productivity. This groundbreaking methodology illustrates Bonsai 27B's dedication to advancing the frontiers of mobile AI technology and enhancing the user experience across diverse applications.

Ling 3.0 Tiny is an advanced reasoning model with open weights, consisting of 7.9 billion parameters in total and 1.3 billion that are active, while boasting a remarkable context window of 262,000 tokens. Utilizing a mixture-of-experts architecture, it expands the open-weights Pareto frontier in intelligence relative to its active parameters, all while maintaining a compact size suitable for deployment in various settings. With a score of 25 on the Artificial Analysis Intelligence Index, it rivals gpt-oss-120b, which has a score of 24, even though it uses 15 times fewer total parameters and 4 times fewer active parameters. This exceptional efficiency in parameters comes with a cost, as it demands a hefty 213 million output tokens to finalize the Intelligence Index evaluation. Moreover, Ling 3.0 Tiny shows significant progress in mitigating hallucination rates when compared to Ling-mini-2.0; it boosts its AA-Omniscience score by an impressive 59 points while maintaining consistent accuracy. Rather than resorting to random guesses in uncertain scenarios, the model opted to attempt only 37% of the posed questions during assessment, which resulted in a drastically lowered hallucination rate of 30%, a substantial improvement from the previous generation's staggering 96%. This strategic decision not only underscores the model's enhanced reasoning abilities but also emphasizes its potential for practical applications in the real world. Overall, Ling 3.0 Tiny exemplifies a significant step forward in the development of efficient and reliable AI models.

CodeGemma is an impressive collection of efficient and adaptable models that can handle a variety of coding tasks, such as middle code completion, code generation, natural language processing, mathematical reasoning, and instruction following. It includes three unique model variants: a 7B pre-trained model intended for code completion and generation using existing code snippets, a fine-tuned 7B version for converting natural language queries into code while following instructions, and a high-performing 2B pre-trained model that completes code at speeds up to twice as fast as its counterparts. Whether you are filling in lines, creating functions, or assembling complete code segments, CodeGemma is designed to assist you in any environment, whether local or utilizing Google Cloud services. With its training grounded in a vast dataset of 500 billion tokens, primarily in English and taken from web sources, mathematics, and programming languages, CodeGemma not only improves the syntactical precision of the code it generates but also guarantees its semantic accuracy, resulting in fewer errors and a more efficient debugging process. Beyond just functionality, this powerful tool consistently adapts and improves, making coding more accessible and streamlined for developers across the globe, thereby fostering a more innovative programming landscape. As the technology advances, users can expect even more enhancements in terms of speed and accuracy.

The o3-mini, developed by OpenAI, is a refined version of the advanced o3 AI model, providing powerful reasoning capabilities in a more compact and accessible design. It excels at breaking down complex instructions into manageable steps, making it especially proficient in areas such as coding, competitive programming, and solving mathematical and scientific problems. Despite its smaller size, this model retains the same high standards of accuracy and logical reasoning found in its larger counterpart, all while requiring fewer computational resources, which is a significant benefit in settings with limited capabilities. Additionally, o3-mini features built-in deliberative alignment, which fosters safe, ethical, and context-aware decision-making processes. Its adaptability renders it an essential tool for developers, researchers, and businesses aiming for an ideal balance of performance and efficiency in their endeavors. As the demand for AI-driven solutions continues to grow, the o3-mini stands out as a crucial asset in this rapidly evolving landscape, offering both innovation and practicality to its users.

The o4-mini model, a refined version of the o3, was engineered to offer enhanced reasoning abilities and improved efficiency. Designed for tasks requiring intricate problem-solving, it stands out for its ability to handle complex challenges with precision. This model offers a streamlined alternative to the o3, delivering similar capabilities while being more resource-efficient. OpenAI's commitment to pushing the boundaries of AI technology is evident in the o4-mini’s performance, making it a valuable tool for a wide range of applications. As part of a broader strategy, the o4-mini serves as an important step in refining OpenAI's portfolio before the release of GPT-5. Its optimized design positions it as a go-to solution for users seeking faster, more intelligent AI models.

Llama, a leading-edge foundational large language model developed by Meta AI, is designed to assist researchers in expanding the frontiers of artificial intelligence research. By offering streamlined yet powerful models like Llama, even those with limited resources can access advanced tools, thereby enhancing inclusivity in this fast-paced and ever-evolving field. The development of more compact foundational models, such as Llama, proves beneficial in the realm of large language models since they require considerably less computational power and resources, which allows for the exploration of novel approaches, validation of existing studies, and examination of potential new applications. These models harness vast amounts of unlabeled data, rendering them particularly effective for fine-tuning across diverse tasks. We are introducing Llama in various sizes, including 7B, 13B, 33B, and 65B parameters, each supported by a comprehensive model card that details our development methodology while maintaining our dedication to Responsible AI practices. By providing these resources, we seek to empower a wider array of researchers to actively participate in and drive forward the developments in the field of AI. Ultimately, our goal is to foster an environment where innovation thrives and collaboration flourishes.

OpenELM is a series of open-source language models developed by Apple. Utilizing a layer-wise scaling method, it successfully allocates parameters throughout the layers of the transformer model, leading to enhanced accuracy compared to other open language models of a comparable scale. The model is trained on publicly available datasets and is recognized for delivering exceptional performance given its size. Moreover, OpenELM signifies a major step forward in the quest for efficient language models within the open-source community, showcasing Apple's commitment to innovation in this field. Its development not only highlights technical advancements but also emphasizes the importance of accessibility in AI research.

LTM-2-mini is designed to manage a context of 100 million tokens, which is roughly equivalent to about 10 million lines of code or approximately 750 full-length novels. This model utilizes a sequence-dimension algorithm that proves to be around 1000 times more economical per decoded token compared to the attention mechanism employed by Llama 3.1 405B when operating within the same 100 million token context window. Additionally, the difference in memory requirements is even more pronounced; running Llama 3.1 405B with a 100 million token context requires an impressive 638 H100 GPUs per user just to sustain a single 100 million token key-value cache. In stark contrast, LTM-2-mini only needs a tiny fraction of the high-bandwidth memory available in one H100 GPU for the equivalent context, showcasing its remarkable efficiency. This significant advantage positions LTM-2-mini as an attractive choice for applications that require extensive context processing while minimizing resource usage. Moreover, the ability to efficiently handle such large contexts opens the door for innovative applications across various fields.

The o3-mini-high model created by OpenAI significantly boosts the reasoning capabilities of artificial intelligence, particularly in deep problem-solving across diverse fields such as programming, mathematics, and complex tasks. It features adaptive thinking time and offers users the choice of different reasoning modes—low, medium, and high—to customize performance according to task difficulty. Notably, it outperforms the o1 series by an impressive 200 Elo points on Codeforces, demonstrating exceptional efficiency at a lower cost while maintaining speed and accuracy in its functions. As a distinguished addition to the o3 lineup, this model not only pushes the boundaries of AI problem-solving but also prioritizes user experience by providing a free tier and enhanced limits for Plus subscribers, which increases accessibility to advanced AI tools. Its innovative architecture makes it a vital resource for individuals aiming to address difficult challenges with greater support and flexibility, ultimately enriching the problem-solving landscape. Furthermore, the user-centric approach ensures that a wide range of users can benefit from its capabilities, making it a versatile solution for different needs.