Large Language Models (LLMs) AI Tools
Open-weight large language models available for local deployment, fine-tuning, and integration into applications.
Open weight models exist so inference can run somewhere other than a vendor API. Teams reach for them under a hard constraint: data that cannot leave a network, per token costs that stop making sense at volume, or fine tuning on proprietary text. Listed here are the weights themselves, not the runtimes that serve them. In 2026 the entries sort into three camps. Sparse mixture of experts models at frontier scale, Llama 4 and DeepSeek-V3, activate a fraction of their parameters per token but still need every parameter resident in memory. Dense mid size models such as Gemma 3 and Phi-4 trade capability for fitting on one accelerator. Reasoning tuned models like DeepSeek-R1 spend far more output tokens to buy accuracy, paying in latency. A reasonable starting point is Gemma 3 at 4B or 12B, or Qwen, both of which run on one consumer card and have the widest tooling support. Frontier scale weights can wait. Check licensing first, because open weights and open source are not the same claim. Llama ships under a community license with usage restrictions and a monthly active user threshold, and Gemma carries its own terms plus a prohibited use policy the vendor can revise. Projects with a real redistribution or audit requirement are safer on OLMo or gpt-oss, both Apache 2.0. The hardware floor pushes the same way: DeepSeek-V3 holds 671B total parameters, still several hundred gigabytes at four bit, so it is a multi GPU commitment no quantization trick removes.