# RWKV

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.

Repository: https://github.com/BlinkDL/RWKV-LM
Canonical: https://ross.abutalabs.com/products/rwkv
Language: Python
License: Apache-2.0
License Family: permissive
Topics: attention-mechanism, deep-learning, gpt, gpt-2, gpt-3, language-model, linear-attention, lstm, pytorch, rnn, transformer, transformers, rwkv, chatgpt
Last push: 2026-08-26T10:29:36+00:00

## Health v2 (maintenance only)
Score: 67/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 8, longevity 100
- inputs: {"age_days": 1851, "days_push": 7, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 14683, forks 1020 (observed 2026-08-28T04:11:07.430760+00:00)

## What it is
RWKV is a novel language model architecture that combines RNN efficiency (linear time, constant memory, no KV-cache) with transformer-level LLM performance and parallelizable GPT-style training. The current RWKV-7 'Goose' release includes model weights, training code, and inference tooling, and is a Linux Foundation AI project under Apache-2.0.

## Use cases
- train a large language model with linear-time attention instead of transformers
- run LLM inference with constant memory and no KV-cache
- deploy a chatbot locally on mobile or desktop
- run LLM inference on GPU with high throughput for large batch sizes
- get free sentence embeddings from a language model
- experiment with RNN-based alternatives to GPT architectures
- run an LLM with effectively infinite context length

## When to choose
- you need constant-memory, linear-time LLM inference without KV-cache growth
- you want to train a GPT-like model but with RNN efficiency at inference
- you need efficient on-device or mobile LLM inference
- you want very high batch-size inference throughput on a single GPU

## When to avoid
- you need the broad ecosystem and tooling compatibility of mainstream transformer models
- your stack depends on fine-tuning frameworks built only for standard attention transformers
- you require maximum community support and prebuilt integrations over architectural efficiency

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, deep-learning, llm-inference, llm-training, transformers, sdk
- domain: large-language-models, deep-learning, machine-learning, artificial-intelligence
- platform: python, cross-platform, windows
- tags: rnn, linear-attention, attention-free, language-model, pytorch, rwkv, gpt, constant-memory, infinite-context, chat-model, natural-language-processing, gpu, linux, macos, docker

## Member repositories
- BlinkDL/RWKV-LM (main) score 67
- BlinkDL/ChatRWKV (frontend) score 73
- BlinkDL/AI-Writer (examples) score 32

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:11:07.430760+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:12:25.416484+00:00, confidence not recorded.
  - readme: https://github.com/BlinkDL/RWKV-LM (fetched 2026-08-28T04:11:07.430760+00:00, sha 372ce135cacd)
- Data as of 2026-08-30T08:39:29.467469+00:00.