An open-source AI model is a large language model whose trained parameters — the numerical “weights” that encode what the model knows and how it responds — are released publicly. Anyone can download, run, and modify them without asking permission or paying per-query fees.

Open-weight vs. truly open-source

The term “open-source” covers a wide spectrum in AI. Most models called open-source are more precisely open-weight: Meta’s Llama family, Google’s Gemma, and Alibaba’s Qwen release their trained parameters but withhold training data details or impose license restrictions. A smaller group — including DeepSeek V3/R1 and Allen AI’s OLMo — release weights, code, and data documentation under truly permissive licenses such as MIT or Apache 2.0, meeting the formal Open Source AI Definition published by the Open Source Initiative in October 2024.

Practically speaking, “open-weight” is already transformative. You can run the model on your own hardware, fine-tune it on your own data, and audit its behavior — even if you cannot reproduce the training process from scratch.

The landscape of open models

A few model families dominate the ecosystem:

  • Meta Llama — the series that sparked the modern open-weight era. Llama 3.3 (70B parameters, December 2024) matches performance once locked behind closed APIs. Meta’s license allows commercial use for companies with under 700 million monthly active users.
  • Mistral 7B / Mixtral — French startup Mistral’s early releases carry an Apache 2.0 license: fully unrestricted for any commercial use. Among the first small models to prove that compact, efficient architectures can punch above their parameter count.
  • DeepSeek R1 — a January 2025 milestone: an MIT-licensed reasoning model that matched OpenAI o1’s benchmark performance, reportedly trained for a fraction of what frontier closed models cost.
  • Alibaba Qwen — by 2025 the most-downloaded model ecosystem on Hugging Face, with sizes from 0.5B to 235B parameters.
  • Microsoft Phi — a series of 1B–14B parameter models demonstrating that training on high-quality synthetic data outperforms simply scaling parameter counts.

Why it matters

Cost. Running a self-hosted open model costs only the compute you use. At high query volumes, this can significantly undercut commercial API pricing.

Privacy. Every prompt sent to a closed API leaves your infrastructure. A self-hosted model keeps sensitive data — medical records, legal documents, internal business data — entirely on your own servers. This is increasingly required by regulators, especially in healthcare and finance.

Customization. Open weights can be fine-tuned on domain-specific data using lightweight techniques like LoRA, transforming a general model into a domain specialist that a closed API rarely matches.

Transparency. Researchers can inspect weights to audit bias, test safety properties, and understand model behavior in ways not possible with proprietary black boxes.

The trade-offs

Open models have real costs. Hardware is the most immediate: a 7B model needs roughly 6–8 GB of GPU memory to run at practical speed; 70B models require high-end server hardware. Infrastructure expertise — setting up inference servers, managing quantization, monitoring performance — adds engineering overhead.

Performance still trails frontier closed models on the hardest tasks. The gap has narrowed dramatically since 2023, but top-tier closed models retain an edge on complex multi-step reasoning and cutting-edge multimodal tasks.

How to start

The lowest-friction entry point is Ollama, a free tool that installs in minutes and lets you pull and run models with a single command:

ollama pull llama3.3
ollama run llama3.3

Ollama handles quantization automatically and runs models as a local service with an OpenAI-compatible API — making it easy to connect to existing tools and scripts. For a graphical interface, LM Studio offers a no-command-line experience on Mac and Windows.

To explore the full model catalog, Hugging Face hosts over two million public models with benchmarks, licensing information, and ready-to-use code examples.

In the news

On June 25, 2026, Liquid AI released LFM2.5-230M — a 230-million-parameter open model running at 213 tokens per second on a smartphone CPU, fast enough for on-device AI agents without any cloud connectivity. The model was trained on 19 trillion tokens — far more than models many times its size — showing that training efficiency matters as much as raw parameter count. Read more: Liquid AI Ships 230M Open Model for On-Device AI Agents.

FAQ

Are open-source AI models free to use commercially?
It depends on the license. Apache 2.0 and MIT models (Mistral 7B, DeepSeek R1, many Qwen variants) carry no commercial restrictions. Meta’s Llama family requires accepting a custom license — free for most businesses, but restricted for companies exceeding 700 million monthly active users.

Do I need a powerful GPU to run them?
A modern laptop with 8–16 GB of RAM can run 3B–7B models at practical speeds. Apple Silicon MacBooks handle these well due to their unified memory architecture. Larger models (30B+) benefit from dedicated GPU hardware.

Are open models less safe than closed models?
Commercial APIs include content filters and safety guardrails by default. Open models lack these out of the box; adding them requires deliberate work. For sensitive deployments, this is worth planning for.

What is quantization?
Quantization compresses model weights to use less memory — for example, representing each weight with 4 bits instead of 16. It reduces hardware requirements significantly with only a small quality trade-off, making it practical to run larger models on consumer hardware.