RWKV (pronounced RWaKuV) is an RNN with GPT-level large language model (LLM) performance that can be trained directly like a GPT Transformer (parallelizable).
RWKV combines the best features of RNN and Transformer: excellent performance, constant memory usage, constant inference generation speed, "infinite" context length, and free sentence embeddings. It is also 100% free of self-attention mechanisms.
The RWKV project was initially proposed by Bo Peng (Blink_DL), and as the project gained attention, it gradually developed into an open-source community.
On September 20, 2023, the RWKV open-source project officially joined the Linux Foundation. Today, the RWKV project is an open-source non-profit organization under the Linux Foundation, with some computing power previously supported by sponsors.
RWKV-7 (Goose) is the latest version of the RWKV architecture. The paper was co-authored by Bo Peng and the RWKV community, published on March 18, 2025.
- RWKV-7 Paper: "RWKV-7 Goose with Expressive Dynamic State Evolution"
- Paper Link: arXiv:2503.14456
RWKV-7 adopts Dynamic State Evolution, surpassing the fundamental limitations of the TC0 expressive power of the attention/linear attention paradigm.
::: details Click to view RWKV-7 Architecture Diagram
:::
RWKV 5/6 (Eagle/Finch) architectures have several improvements based on the RWKV-4 architecture. Therefore, these two architectures are published in the same paper.
- RWKV 5/6 Paper: "Eagle and Finch: RWKV with Matrix-Valued States and Dynamic Recurrence"
- Paper Link: arXiv:2404.05892
RWKV-4 is the first official version of the RWKV model. The paper was co-authored by Bo Peng and the RWKV community and was first published on May 22, 2023. In October of the same year, the RWKV-4 architecture paper was accepted by EMNLP 2023.
- RWKV-4 Paper: "RWKV: Reinventing RNNs for the Transformer Era"
- Paper Link: arXiv:2305.13048
RWKV has released open-source models of various parameter scales for each architecture version.
| Version | RWKV-V4 | RWKV-v5-Eagle | RWKV-v6-Finch | RWKV-v7-Goose | RWKV-v7-G1 |
|---|---|---|---|---|---|
| Paper | Published | Published | Published | Published | Coming Soon |
| Overall Status | EOL | EOL | Stable | Stable | In Progress |
| 0.4B Model | Released | Released | No Plan | Released | Released |
| 1.5B Model | Released | Released | Released | Released | 📅 Planned |
| 3B Model | Released | Released | Released | Released | 📅 Planned |
| 7B Model | Released | Released | Released | 📅 Planned | 📅 Planned |
| 14B Model | Released | No Plan | Released | 📅 Planned | 📅 Planned |
Please use RWKV-7 series models. RWKV-7 models are based on the latest RWKV-7 architecture and latest datasets, therefore offering better performance.
Since RWKV-7 7B and larger models are still in training for 7B and larger parameter models, please use the RWKV-6-World-14B-V2.1 model; consider using the RWKV-6-World-7B-V3 model if your hardware cannot run the 14B model.
::: tip RWKV-7-World 7B/14B will replace the existing RWKV-6-World 7B/14B models once training is complete. Earlier RWKV versions have come to the end of their lifecycle, and existing models are only for archival purposes. :::
-
Advantages
- Lower resource usage during runtime and training (VRAM, CPU, GPU, etc.).
- 10 to 100 times lower computational requirements compared to Transformers with larger contexts.
- Supports linear scaling to any context length (Transformers scale quadratically).
- Performs as well as Transformer architectures in terms of answer quality and generalization ability.
- RWKV models' training data includes languages other than English (e.g., Chinese, Japanese, etc.), offering better multilingual capabilities than most existing open-source models.
-
Disadvantages
- RWKV base models are very sensitive to the format of prompts, and the format of prompts significantly affects the generation results.
- Due to architectural design, RWKV models are weaker in tasks requiring retrospection, so prompts need to be appropriately ordered. For example, provide task instructions to the model first, then provide the material text needed to perform the task.
| Concept | Description |
|---|---|
| RWKV | The model architecture itself, training code can be found here. |
| ChatRWKV | The official chatbot of RWKV (similar to ChatGPT but based on RWKV), code can be found here. |
| RWKV-4/5/6/7 | Different architecture versions of RWKV. Note that using the latest RWKV-7 series models is recommended. |
| RWKV World | The base RWKV model trained with global languages, covering a broader and more diverse dataset, including training data in over 100 languages and some instruction training. |
| Raven | The official fine-tuned version of the RWKV-4 base model, including instruction training. However, since the RWKV-4 series is no longer updated, it is not recommended for continued use. |
| RWKV ABC/MIDI | RWKV music models based on ABC/MIDI format |
| RWKV CHNtuned / one-state-chat / role_play / novel ... | Fine-tuned models provided by the RWKV community, optimized for specific tasks or data types. Please prioritize using RWKV-7 series fine-tuned models. |
RWKV models typically have two naming conventions:
- RWKV-6-World-3B-v2.1-20240208-ctx4096.pth
- RWKV-x070-World-1.5B-v3-20250127-ctx4096.pth
The meaning of each field in the model name:
| Field | Meaning |
|---|---|
| RWKV | Model name |
| 6 / 070 | RWKV model architecture, recommended to use RWKV-7 models |
| World | Model type, World indicates RWKV models trained with global languages, thus supporting multilingual tasks |
| 3B / 1.5B | Model parameter scale, "B" stands for "Billions" |
| v2.1 / v3 | Model training dataset version, v2.1 ≈ 1.1T , v3 ≈ 2.5T |
| 20240208 / 20250127 | Model release date |
| ctx4096 | Pre-trained context length |
| .pth | RWKV model file format, also supports .gguf and .safetensors etc. |
RWKV is made possible, as an Open Source project, thanks to the large amount of GPU compute and researchers time contributions from
Without their invaluable support, we would not have been able to develop the core RWKV foundation models that you see today.
In addition, we would like to thank
- alpin @ pygmalionAI
- AutoMeta @ AlignmentLab
- FeatherlessAI
- Various other folks who donated slices of GPU time / preferred not to be named
For helping with GPU time, on smaller experiments, finetunes, and various models. Especially for those models that never get publically released in failed runs.