RWKV-7 "Goose" Performance Summary: RWKV Models as of Early 2026

Forum topic · ✨步子哥 · 2026-02-12

Summary

A Chinese forum post summarizes the performance of RWKV-7 "Goose" models as of early 2026. RWKV is a pure RNN architecture with no attention mechanism and linear inference cost. Key claims: inference time is linear in sequence length, memory usage is constant (no KV cache), training is parallelizable like a Transformer, and models run efficiently on phones and integrated GPUs. Benchmark results are given for four sizes: 13.3B (G0b) achieves MMLU 76.5%, GSM8K 92.3%, MATH 76.8%, IFEval 68.9%, and the best Uncheatable Eval score of 6.843 (lower is better, reported better than Qwen3-14B); 7.2B, 2.9B, and 1.5B variants are also tabulated. Measured inference speeds for the 2.9B model include 115 tokens/s on an RTX 4090 (nf4, 2.4 GB VRAM), 86 t/s on an RTX 4060 Laptop, 30+ t/s on a Snapdragon 8 Gen 3 phone, and 6.5 t/s on an RK3588 NPU, with claimed 3-10x speed and 1/3 memory versus Transformers. Recommendations include web-rwkv or llama.cpp backends and World-series models for multilingual tasks.

This post from zhichai.net presents a performance summary of the RWKV-7 "Goose" model series as of early 2026. RWKV is a pure RNN architecture with no attention mechanism, offering linear-time inference.

Core Advantages

Linear inference time: no quadratic complexity bottleneck; runtime is independent of sequence length.
Constant memory: no KV cache; very low VRAM usage, supporting unlimited context length.
Parallelizable training: trains as efficiently as a Transformer, breaking the serial limitation of classic RNNs.
Highly efficient: runs in real time on phones/integrated GPUs with large power savings.

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 |

*Note: lower is better for Uncheatable Eval; the 13.3B model is reported to outperform Qwen3-14B on this benchmark.*

Measured Inference Performance (RWKV-7 2.9B)

RTX 4090: 115 tokens/s (nf4 quantization, 2.4 GB VRAM)
RTX 4060 Laptop: 86 tokens/s (nf4, 2.4 GB VRAM)
Snapdragon 8 Gen 3 phone: 30+ tokens/s (W4A16 quantization, edge-usable)
RK3588 NPU: 6.5 tokens/s (W8A8, embedded devices)

Compared with Transformers, the post claims 3-10x faster speed and only ~1/3 of the memory usage.

Usage Recommendations

Maximum performance: choose RWKV-7 13.3B / 7.2B, approaching or exceeding mainstream Transformer models.
Phone/laptop deployment: the 2.9B G1 series (GGUF) runs smoothly on ordinary hardware.
Multilingual tasks: prefer the World series, reported at SOTA level on multilingual benchmarks.
Recommended backends: web-rwkv (fastest), llama.cpp (general-purpose).

Resources

Website: www.rwkv.com
Wiki: wiki.rwkv.com
Models: Hugging Face - BlinkDL/rwkv7-g1