Open-Source AI

Open-source AI refers to AI models, frameworks, or systems whose weights, code, or architecture are publicly released — allowing anyone to download, inspect, modify, and deploy them — in contrast to closed proprietary models accessible only through a paid API.

Open-source AI encompasses a broad range of public releases in the AI domain, from deep learning frameworks (PyTorch, TensorFlow) and training datasets (Common Crawl, The Pile) to the full weights of large language models. The most consequential open releases have been foundation model weights: Meta's LLaMA series (LLaMA 1 in February 2023, LLaMA 3 in April 2024, and subsequent versions), Mistral 7B and Mixtral from Mistral AI (late 2023), and DeepSeek-R1 (DeepSeek, January 2025), which demonstrated reasoning capabilities broadly competitive with GPT-4o at dramatically lower reported training cost, sending significant shockwaves through AI industry valuations.

The definition of "open source" in AI is actively contested. The Open Source Initiative released a formal Open Source AI Definition in October 2024, requiring release of sufficient information — including training data, code, and weights — under terms permitting use, study, modification, and redistribution without restriction. Most high-profile model releases fall short of this standard: Meta's LLaMA releases provide weights and inference code but not training data; many add commercial restrictions (e.g., prohibiting use above certain revenue thresholds). The industry term "open weights" is therefore increasingly used to distinguish models where only parameters are public from software that is open source in the traditional sense.

Open-weight AI matters for several interconnected reasons. It enables researchers to study model internals, identify failure modes, and reproduce results — a prerequisite for scientific rigor that closed APIs preclude. It allows companies and governments to self-host models without incurring per-token API fees or routing data through third parties, which is critical for privacy-sensitive or regulated applications. It also accelerates ecosystem development: thousands of fine-tuned variants, evaluation benchmarks, and serving frameworks (llama.cpp for CPU inference, vLLM for throughput-optimized GPU serving, Ollama for local deployment) have been built on open-weight foundations, creating a parallel innovation track alongside closed-source labs.

As of 2026, the capability gap between the best open-weight models and frontier closed-source models has narrowed substantially. DeepSeek-V3 and its successors showed that highly capable models can be trained at reported costs far below those of OpenAI or Anthropic flagships, challenging assumptions that compute scale alone would keep closed labs permanently ahead. Regulatory debates — particularly under the EU AI Act — center on whether highly capable open-weight models warrant additional obligations, since once weights are public they cannot be retracted. Proponents argue openness produces net social benefit through transparency and access; some AI safety researchers argue it removes a meaningful access-control layer for potentially dangerous capabilities.