ChatRWKV is an open-source chatbot built on the RWKV language model architecture, offering a ChatGPT-style experience from a model that is 100 percent a recurrent neural network rather than a Transformer. The GitHub project has attracted more than 9,400 stars and continues to receive updates in 2026.
RWKV blends the parallelizable training of Transformers with the efficient, constant-memory inference of RNNs. ChatRWKV wraps that architecture in a usable chat program: it loads RWKV checkpoints of many sizes, supports strategies that trade speed against memory across GPUs and CPUs, and can be fine-tuned on chat data to follow instructions. Because inference state stays compact, RWKV models can generate very long outputs without the growing key-value cache that makes Transformer inference expensive, which makes the project a fascinating working proof of an alternative path for language models. The repository also documents the training and fine-tuning pipelines behind the released models, so it is possible to reproduce the whole journey from a pretrained base model to a working chat assistant.
ChatRWKV is free and open source under the Apache License 2.0, and the RWKV model weights are freely available. Your only cost is the hardware you run it on.
ChatRWKV appeals to researchers and hobbyists curious about post-Transformer architectures, and to developers who need efficient long-context generation on constrained hardware. It is also a hands-on way to evaluate whether RNN-based models fit workloads that Transformers serve today.