RWKV-7 "Goose" Performance Summary (Early 2026)
A Chinese forum post presents a summary poster of RWKV-7 "Goose" model performance as of early 2026. RWKV-7 is a pure RNN architecture with no attention mechanism and linear inference cost.
Key Architectural Advantages
- Linear inference time: no quadratic complexity bottleneck; cost is independent of sequence length.
- Constant memory: no KV cache, very low VRAM usage, unlimited context support.
- Parallelizable training: trains as efficiently as a Transformer, breaking traditional RNN serialization limits.
- Highly efficient: runs in real time on phones / integrated GPUs with low power consumption.
- Maximum performance: choose RWKV-7 13.3B / 7.2B, approaching or exceeding mainstream Transformer models.
- Phone/laptop deployment: recommended 2.9B G1 series (GGUF), runs smoothly on ordinary hardware.
- Multilingual tasks: prefer the World series, claimed SOTA on multilingual benchmarks.
- Recommended backends:
web-rwkv(fastest),llama.cpp(general-purpose). - Website: www.rwkv.com
- Wiki: wiki.rwkv.com
- Hugging Face: BlinkDL/rwkv7-g1
RWKV-7 Benchmark Results
| Model Size | MMLU | GSM8K | MATH | IFEval | Uncheatable Eval | |---|---|---|---|---|---| | 13.3B (G0b) | 76.5% | 92.3% | 76.8% | 68.9% | 6.843 (Best) | | 7.2B (G0a3) | 65.1% | 83.9% | 67.8% | 58.0% | 7.222 | | 2.9B (G1a4) | 61.3% | 77.3% | 48.2% | 51.0% | 7.486 | | 1.5B (G1b) | 50.5% | 58.5% | 29.8% | 42.1% | 7.969 |
*Lower is better for Uncheatable Eval; the 13.3B model reportedly outperforms Qwen3-14B on this benchmark.*
Real-World Inference Performance (RWKV-7 2.9B)
| Hardware | Speed | Notes | |---|---|---| | RTX 4090 | 115 t/s | nf4 quantization, 2.4 GB VRAM | | RTX 4060 Laptop | 86 t/s | nf4 quantization, 2.4 GB VRAM | | Snapdragon 8 Gen 3 (phone) | 30+ t/s | A16W4 quantization, edge-usable | | RK3588 NPU | 6.5 t/s | W8A8, embedded device |
Claimed vs. Transformer: 3-10x faster, ~1/3 the memory.