BlinkDL/RWKV-LM
RWKV (pronounced RwaKuv) is an RNN with great LLM performance, which can also be directly trained like a GPT transformer (parallelizable). We are at RWKV-7 "Goose". So it's combining the best of RNN and transformer - great performance, linear time, constant space (no kv-cache), fast training, infinite ctx_len, and free sentence embedding.
Implements an attention-free RNN architecture with state-passing mechanisms that enable meta-in-context learning—adapting model state via gradient descent at inference time. RWKV-7 achieves linear-time complexity with zero KV-cache overhead while maintaining training parallelizability through careful initialization (PreLN LayerNorm), selective weight decay on projection matrices only, and differentiated learning rates per parameter. Offers production-ready inference via CUDA kernels (206k tokens/s on H100 clusters), mobile libraries, and Windows/Office runtime integration, with reference training code requiring only 10GB VRAM on a single GPU.
14,414 stars. Actively maintained with 2 commits in the last 30 days.
Stars
14,414
Forks
997
Language
Python
License
Apache-2.0
Category
Last pushed
Mar 05, 2026
Commits (30d)
2
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/transformers/BlinkDL/RWKV-LM"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.
Related models
intel/auto-round
🎯An accuracy-first, highly efficient quantization toolkit for LLMs, designed to minimize quality...
ModelCloud/GPTQModel
LLM model quantization (compression) toolkit with hw acceleration support for Nvidia CUDA, AMD...
pytorch/ao
PyTorch native quantization and sparsity for training and inference
Picovoice/picollm
On-device LLM Inference Powered by X-Bit Quantization
NVIDIA/kvpress
LLM KV cache compression made easy