RWKV-LM is the home of RWKV, an alternative neural network architecture that delivers transformer-level language model performance with the efficiency of an RNN. Pronounced “RwaKuv”, RWKV has grown from a solo research project into a Linux Foundation AI project with a large research community, and its GitHub repository has earned more than 14,000 stars, with the current RWKV-7 generation under active development.
Transformers pay for their power with quadratic attention costs and a growing KV cache. RWKV takes a different route: it is 100% recurrent, so inference runs in linear time with constant memory, no KV cache, and effectively unlimited context length, yet it can be trained in parallel just like a GPT model. The latest generation, RWKV-7 “Goose”, acts as a meta in-context learner, continually adapting its state to the context as it generates. The repository contains the reference implementations, training code, and model tooling across generations, plus links to hundreds of papers that have trained RWKV variants for language, multimodal, and other tasks. RWKV runtimes are already integrated into products and platforms, including parts of the Windows ecosystem, and community ports cover mobile and embedded deployment.
RWKV is completely free and open source under the Apache 2.0 license, with open model weights released by the community. There are no fees, tiers, or restrictions beyond the license itself.
RWKV-LM is for researchers exploring post-transformer architectures, engineers who need long-context inference on limited hardware, and builders who want LLM capability without attention’s memory costs. If you are curious where sequence modeling goes beyond transformers, this is one of the most credible open answers.