Picture by the writer
. Introduction
Small language models (SLM) are rapidly becoming the practical face of AI. They are getting faster, smart and much more efficient, which provides strong results with a part of computers, memory, and energy that require major models.
The AI ​​community is to use a large language model (LLM) to develop artificial datases in the growing trend, which is then used to fix SLMs for specific tasks or adopt a special style. As a result, the SLMs are smart, fast and more expertise, maintaining a compact size. This opens up interesting possibilities: Now you can now embed the intelligent models directly into systems that do not require permanent internet connection, which can enable the On -On -On -On -On -On -On -On -On -On -on -on -on -on -on -on -on -on -on – -that
In this tutorial, we will review the small models making waves in the AI ​​world. We will compare their size and performance, which will help you understand which models offer the best balance for your needs.
. 1. Google/JEMA-3-270m -T
JEMA 3 270 m Model Jima 3 is the smallest and most extreme ultra -lightweight member of the family, designed for performance and access. With only 270 million parameters, it can easily operate on limited computational resource devices, which is ideal for experience, prototype, and lightweight applications.
Despite its compact size, the 270 -meter model supports the 32K context window and can handle broad tasks such as answering basic questions, summary and reasoning.
. 2. Qwen/qwen3-0.6b
Qwen3-0.6b Model Qian 3 is the lightest variety of series, designed to provide strong performance during the most efficient and accessible residual. With 600 million parameters (0.44b non -embedding), it balances the capacity and resource needs.
QWen3-0.6B comes with the ability to switch without interruptions between “thinking mode” for complex reasoning, mathematics, and coding, and “non-thinking mode” for sharp, general purposes. It supports the length of 32K context and offers multilateral support in 100+ languages.
. 3. The hugs work
smollm3-3b The model is a small powerful open source language model designed to advance the boundaries of small -scale language models. With 3 billion parameters, it offers strong performance in arguments, mathematics, coding, and multi -linguistic tasks, while the broader access to enlighten is quite efficient.
The SMALLM3 supports the dual -mode reasoning, which allows users to solve the complex problem for general dialogue and toggle between the “mood of thinking” of a sharp, lightweight mode.
Beyond the text generation, the SMollM3 also enables the use of an agent with tool calling, which makes it a versatile of real -world applications. Public training details, as a fully open model with open weight and checkpoints, provides researchers and developers a transparent, high performance base for creating a 3B-4B scale reasoning on a 3B-4B scale.
. 4.
QWEN3-4B-Instruct-25507 The latest instruction of the model QWen3-4B series is different format, designed to provide strong performance in non-thinking mode. With 4 billion parameters (3.6B non -disconnected), it also introduces significant improvement in logical reasoning, understanding of text, mathematics, science, coding and device use in several languages.
Unlike other QWen3 models, this version has been specially improved for non -thinking mode, which ensures a faster, more efficient reaction without producing the reasoning token. It also demonstrates better alignment with user preferences, which perform well in open and creative tasks such as writing, dialogue and saplus arguments.
. 5. Google/JEMA -3-4 B. These
JEMA 3 4 B The model is a multi -model member of a direction, JEMA 3 Family, designed to handle both text and image input while producing high quality text output. With 4 billion parameters and 128K token context window, it is suitable for tasks such as answers to questions, summary, reasoning, and detailed image understanding.
The important thing is that it is highly used for text rating, image rating, or fine toning on special tasks, which improves model skills and performance for some domains.
. 6. Joe Q/Jan-V 1-4 B
Jan-V 1 The Model John Family is the first release, especially designed to solve the agent reasoning and the problem within the John app. On the basis of the Lucy Model and powered by the QWen3-4B-thinking architecture, John-V-One offered better reasoning capabilities, tool use and better performance on complex agent tasks.
By scaling the model and fixing its parameters, it has achieved an impressive accuracy of 91.1 % on the simple cue. In a realistic question that responds to models of this size, it indicates an important milestone. With the recommended settings to enhance efficiency, local use with John app, VLM, and Llama.Cpp has improved it.
. 7. Microsoft/Phi-4-Mini-Instruct
Phi-4-Mini-Instruct The model is Microsoft’s PHI4 family’s lightweight 3.8B parameter language model, which is designed for effective reasoning, instructions, and secure deployment in both research and commercial applications.
Trained at a mixture of 5T tokens from high quality filtered web data, artificial “textbook” reasoning data, and curatic -monitoring instructions data, it supports 128k token context and Excel in mathematics, logic and multi -linguistic works.
The PHI-4-mini-instruct also supports integration with framework such as calling, multi-linguistic breeds (20+ languages), and VLM and transformers, which can enable flexible deployment.
. Conclusion
This article detects a new wave of lightweight open models so far, which is renewing AI landscape by balanced performance, reasoning and access to the landscape.
From Google’s Jima 3 Family with Ultra Compact gemma-3-270m-it And multi -modal gemma-3-4b-itIn the QWEN3 Series of Kevin with Effective Qwen3-0.6B And long context, instructions were improved Qwen3-4B-Instruct-2507These models highlight how scaling and fine toning can unlock strong reasoning and multi -linguistic abilities in small feet marks.
SmolLM3-3B Double -mode leads to small model limits with the help of reasoning and long context, while Jan-v1-4B John app focus on agent reasoning and device use within the ecosystem.
Finally, Microsoft Phi-4-mini-instruct This shows how 3.8B parameters can provide competitive performance in mathematics, logic and multi -linguistic tasks through high quality artificial data and alignment techniques.
Abid Ali Owan For,,,,,,,,,, for,, for,,,, for,,,, for,,, for,,,, for,,,, for,,,, for,,, for,,, for,,, for,,, for,,,, for,,, for,,, for,,,, for,,, for,,,, for,,, for,,, for,,,, for,,, for,,, for,,,, for,,, for,,,, for,,, for,,,, for,,, for,,,, for,,, for,,,, for,,,, for,,,, for,,,, for,,,, for,,,, for,,,, for,,, for,,, for,,, for,,, for,,,,, for,,,, for,,,, for,,,, for,, for,.@1abidaliawan) A certified data scientist is a professional who loves to create a machine learning model. Currently, he is focusing on creating content and writing technical blogs on machine learning and data science technologies. Abid has a master’s degree in technology management and a bachelor’s degree in telecommunications engineering. Its vision is to create AI products using a graph neural network for students with mental illness.