Definition
Open source LLMs are large language models whose trained weights are published for anyone to download, run, inspect and fine-tune on their own hardware. Most models called open source are more precisely open-weight: the parameters are released, but the training data and full training code are not.
The distinction has a formal basis. The Open Source Initiative published the Open Source AI Definition 1.0 in October 2024. It requires detailed information about training data, the complete code used to train and run the system, and the parameters, all under OSI-approved terms. By that standard, very few popular models qualify. OpenAI's gpt-oss-120b and gpt-oss-20b, released in August 2025 under Apache 2.0, are open-weight. So are Meta's Llama models, which ship under Meta's own community license, and Moonshot AI's Kimi K3, released in July 2026 under a custom license that requires a separate agreement for large model-as-a-service businesses and visible attribution in products with more than 100 million monthly users. Always read the license before building a product on a model.
The field shifted toward Chinese labs during 2025 and 2026. DeepSeek-R1, released in January 2025 with open weights, showed that strong reasoning did not require a closed model, and DeepSeek has since moved to its V4 series. In October 2026, independent leaderboards placed Moonshot's Kimi K3 (a 2.8 trillion parameter model with a 1 million token context window), Zhipu's GLM-5 series, Alibaba's Qwen3.8 and Xiaomi's MiMo-V2.6 among the strongest open-weight models, with Mistral Large 4 and Google's Gemma 4 family also widely used. Meta went the other way: after Llama 4 disappointed developers in April 2025, its April 2026 Muse Spark model, which now powers Meta AI, was released as a closed model. Leaderboard positions change monthly and depend heavily on evaluation settings, so treat any ranking as a snapshot.
Teams choose open-weight models for control. Data can stay on their own infrastructure, costs scale with hardware rather than per-token fees, and the model can be fine-tuned on domain data or distilled into small language models. The costs are operational: serving large models needs serious GPU capacity, safety tuning and updates become the deployer's job, and the largest open models still trail the best closed models on some agentic and long-horizon tasks. Tools such as vLLM, llama.cpp and Ollama handle much of the serving work, and gpt-oss-20b, for example, runs in 16 GB of memory.
For AI visibility, open-weight models matter mainly through what they already know. A self-hosted model inside a support bot, internal assistant or third-party app usually answers from parametric knowledge frozen at its training cutoff, and it cites nothing unless the developer adds retrieval. How a brand is described across the public web before that cutoff, including documentation, reviews, forums and news, shapes those answers for as long as the model stays deployed. Nobody can monitor every private deployment, but querying the major open models directly through their APIs shows how a brand is represented in their training data. Promptwatch includes models such as DeepSeek and Mistral for this reason: they show what the model itself believes, independent of live search results.
Examples of Open Source LLMs
- A bank fine-tunes gpt-oss-120b on its internal policy manuals and runs it on its own 80 GB GPUs, so customer data never leaves its data center and the Apache 2.0 license allows commercial use.
- A startup evaluating Kimi K3 for a hosted API product reads the license and finds that, because its parent company passes $20 million in annual revenue, it needs a separate agreement with Moonshot AI before launch.
- A software brand queries DeepSeek and Mistral through their APIs and finds both still describe a product line it discontinued in 2024, because their training data predates the change. It updates documentation and third-party listings ahead of the next model releases.
- A research team picks a mid-sized Qwen3.8 model over a larger one after testing both on its own extraction task, since leaderboard gaps between them disappeared under its prompt format.
