Proprietary models aren’t the only ones grabbing attention: for certain uses, configurability matters more than raw power. This quest for control has fueled the rise of locally hosted models. With suitable hardware and tools like Ollama or LM Studio, capable models can run entirely on the user’s device.
- Local models gain value when configurability matters more than raw power.
- Five models fit different uses, and the right choice depends on the task and the hardware.
- Ollama offers the simplest installation and exposes a local API for your own applications.
- Local execution does not guarantee confidentiality, because connected tools and telemetry can still send data elsewhere.
Which models are worth installing, and which ones would run on a given device? The article highlights five for 2026, LLMs (large language models) chosen based on use case and available hardware.
Five models for different uses
The first, Llama 3.1 8B Instruct, serves as a generalist reference for local chat, summarization, classification, and fine-tuning projects, meaning adaptation to specific data. It only handles text and is distributed under Meta’s Llama Community License, the terms of which must be checked before any redistribution or deployment.
For Qwen2.5 7B, the choice depends on the task. The Qwen2.5-7B-Instruct version is suitable for multilingual conversations, document processing, and structured outputs like JSON format. Qwen2.5-Coder-7B-Instruct focuses on generating, explaining, and correcting code. Both are text models: to analyze screenshots, an alternative capable of processing images is needed.
Mistral 7B Instruct v0.3 is a seven-billion-parameter text model, intended for rewriting, summarization, documentation, and everyday dialogue. Its v0.3 version handles function calling, allowing integration with tools executed by an application, and its Apache 2.0 license facilitates customization projects. An older generalist, it’s worth comparing to recent models on its own queries before being selected.
Gemma 4 E2B and E4B accept text, images, and audio, and generate text. They are suitable for questions about documents, image understanding, transcription, or embedded assistance, provided the runtime environment supports the input type. Their names refer to effective parameters: E2B has about 5.1 billion parameters, and E4B has 8 billion, including embeddings. The article warns: hardware sizing should not be based solely on the names.
Phi-4-mini-instruct, from Microsoft, is a 3.8-billion-parameter model aimed at memory-constrained and latency-sensitive applications. It is suitable for small local assistants, information extraction, math or logic problems, and function calling experiments. The reasoning and multimodal models with similar names are separate checkpoints, with different behaviors and prerequisites.
The ranking of model families is based on the cumulative number of downloads in the Ollama library. It therefore favors older versions and does not measure quality.
Installing a model with Ollama
According to the article, Ollama is arguably the simplest option for developers: just install it, download a model, and run it from the terminal, for example with the command ollama run gpt-oss:20b. The tool offers local packages for phi-4, Gemma 4, Qwen3-Coder, and other models. It also exposes a local API, useful for connecting a model to your own applications or programming agents.
The article’s FAQ provides several clarifications. Qwen2.5-Coder-7B-Instruct is presented as the code-oriented option, and generated code should be tested before use. Local execution alone does not guarantee confidentiality: connected tools, cloud fallback solutions, and application telemetry can transmit data elsewhere. A model’s popularity reflects its adoption, but the task and hardware determine the most useful choice.
