RWKV-7 G1j
#21 v Open-source jazykové modelyRwkv · v7 · g1j · 5× · naposledy 03. 9. 2026
40
Momentum
RWKV-7 G1 (Codename "GooseOne") is a series of open, Open Source language models from the RWKV Project (Linux Foundation AI & Data), based on the RWKV-7 "Goose" architecture—a pure RNN without Self-Attention and without KV-Cache. The G1 series enhances the base architecture with significantly more training data and "thinking"/Reasoning capabilities (Chain-of-Thought) and is available in sizes ranging from 0.1B to 13.3B parameters. The models are released under the Apache-2.0 license and run via Hugging Face, Ollama, and llama.cpp/GGUF, among others; the first release version launched in March 2025, with improved data versions (G1a through G1f and beyond) released regularly since.
Vývoj momenta
08.06.06.09.
Vlastnosti
| Key Benchmark (%) | MMLU 61.2% (rwkv7-g1a-2.9b, single-token eval); 2.9B model achieves 71.5% average English benchmark accuracy |
| Context Window (Tokens) | Trained with ctxlen 4096 (e.g. rwkv7-g1-1.5b-ctx4096); as an RNN theoretically unbounded, but reliable recall up to roughly 8k–16k tokens in practice |
| License | Apache 2.0 |
| Multimodality | Text-only (text generation), multilingual (incl. English, Chinese, Japanese, Korean, French, Arabic, Spanish, Portuguese) |
| Platform | Hugging Face, Ollama, llama.cpp/GGUF, RWKV pip package (≥0.8.28) |
| Release Date | First released March 2025 (RWKV7-G1 “GooseOne”), ongoing updates since (G1a–G1f, G1d in Feb 2026, G1e in March 2026) |