RWKV Project

weights

open(Apache-2.0)

data

components-listed(itemized machine-readable component index with resolvable links to public datasets, but no released corpus and no reconstruction pipeline)

code

open(github.com/BlinkDL/RWKV-LM)

license

Apache-2.0(OSI)

RWKV publishes Apache-2.0 weights and open training code, plus a machine-readable index naming every training-data component with a resolvable link, category and citation. But the corpus itself is not released, mixing proportions are withheld, one component is gated and another's version unresolved, so the data stays itemized but not reproducible -- open weights rather than fully open.

• https://api.github.com/repos/BlinkDL/RWKV-LM recorded 2026-08-13

  `private: false`, license `Apache-2.0`, 14,642 stars, pushed 2026-07-23. Model and training code by version (RWKV-v1..v8). No data-prep or corpus-reconstruction directory.

• https://arxiv.org/abs/2503.14456 recorded 2026-09-29

  The RWKV-7 Goose abstract says "we release our models and dataset component listing at" the RWKV Hugging Face org "and our training and inference code at" github.com/RWKV/RWKV-LM, "all under the Apache 2.0 License", and describes "an extended open source 3.1 trillion token multilingual corpus".

• https://huggingface.co/api/datasets/Goose-World/RWKV-World-v3 recorded 2026-09-29

  The dataset API describes the repo as "an itemised and annotated list of the RWKV World v3 corpus" tagged "license:apache-2.0"; gated and private are both false, and the files are only index/train.jsonl, train.csv, the 100k and 1m subsample JSONL files, two READMEs and one figure, with no corpus shards.

• https://huggingface.co/api/models/BlinkDL/rwkv-7-world recorded 2026-09-29

  The model API for BlinkDL/rwkv-7-world is tagged "license:apache-2.0" with gated and private both false, and lists downloadable weight files including "RWKV-x070-World-2.9B-v3-20250211-ctx4096.pth" and "RWKV-x070-World-1.5B-v3-20250127-ctx4096.pth".

• https://huggingface.co/api/models/BlinkDL/rwkv7-g1 recorded 2026-09-29

  The model API for BlinkDL/rwkv7-g1, the line the RWKV-7 World card now points users to, is tagged "license:apache-2.0" with gated and private both false, and lists weight files up to "rwkv7-g1j-13.3b-20260831-ctx16384.pth".

• https://huggingface.co/BlinkDL/rwkv7-g1/raw/main/README.md recorded 2026-09-29

  The RWKV7-G1 card says "These are BASE models (pretrained with web/code/synthetic + instruction/chat/reasoning data)" and points to "Training: https://github.com/BlinkDL/RWKV-LM" for training code; it gives no itemized listing of the added synthetic and instruction data.

• https://huggingface.co/datasets/RWKV/RWKV-World-Listing/raw/main/train.csv recorded 2026-09-29

  The component index opens with the header "Dataset;Category;Categories;Citation;Version;URL;Notes" and gives each component a resolvable URL, but defers sampling to unpublished material ("see wiki.txt in the supplementary material for exact amounts"), flags OSCAR as "gated" and leaves peS2o unresolved ("Which version out of the 3?").

Adoption is measured as Hugging Face downloads summed across the two repositories RWKV declares, rather than GitHub stars, which had overstated the family's reach. Counting every repository under the BlinkDL and RWKV organizations would read higher, but only the artifacts a product declares count toward it.

• https://huggingface.co/api/models/BlinkDL/rwkv-7-world recorded 2026-08-12

  4,395 downloads in the trailing 30 days for BlinkDL/rwkv-7-world

• https://huggingface.co/api/models/RWKV/RWKV7-Goose-World3-1.5B-HF recorded 2026-08-12

  257 downloads in the trailing 30 days for RWKV/RWKV7-Goose-World3-1.5B-HF

RWKV's case is efficiency rather than absolute capability: a linear-time recurrent architecture that its paper shows competitive at 3B scale on far fewer training tokens. Its best published MMLU sits well below InternLM2.5-7B and Yi-34B, small and mid-size models from the 2023-2024 generation that this map places at the same level, and the G1 models that extend the line to 13.3B report only language-modeling results, which give no grounds for placing it higher.

• https://arxiv.org/abs/2503.14456 recorded 2026-10-08

  The abstract says the "2.9 billion parameter language model achieves a new 3B SoTA on multilingual tasks and matches the current 3B SoTA on English language downstream performance", trained "on dramatically fewer tokens than other top models".

• https://huggingface.co/BlinkDL/rwkv-7-world/raw/main/README.md recorded 2026-10-08

  The RWKV-7 World card reports "rwkv7-v3-2.9b 54.56% (rwkv6-v2.1-3.1b 32.38%)" and "rwkv7-v3-1.5b 44.84% (rwkv6-v2.1-1.6b 26.34%)" on its MMLU eval, and tells users to "Please upgrade to https://huggingface.co/BlinkDL/rwkv7-g1" as better in all aspects.

• https://huggingface.co/BlinkDL/rwkv7-g1/raw/main/README.md recorded 2026-10-08

  The RWKV7-G1 card lists sizes up to "13.3B = L61-D4096" and points to UncheatableEval rather than publishing a benchmark table ("check https://huggingface.co/spaces/Jellyfish042/UncheatableEval to see their performance at language modeling").