RWKV combines Transformer-style parallelizable training with recurrent inference. RWKV-APP turns that architecture into real products: curated model packages, hardware-aware runtimes, and one open-source client for phones and desktops.

RWKV 将可并行训练与循环推理结合。RWKV-APP 负责把这种架构变成真实产品：适配模型、打磨硬件运行时，并提供覆盖手机与桌面的开源客户端。

• Parallelizable training / 可并行训练 — time-parallel training with efficient recurrent decoding.
  训练可在时间维并行，生成时采用高效的循环解码。

• Linear-time sequence processing / 线性处理序列 — across a sequence of length T, recurrent inference performs O(T) work; each new token does not revisit the full history.
  对于长度为 T 的序列，循环推理的总计算量为 O(T)，生成新 token 时无需重新处理完整历史。

• Fixed-size recurrent state / 固定大小循环状态 — for a fixed model and one stream, state memory is O(1) with respect to context length.
  给定模型与单条推理流，状态占用相对于上下文长度为 O(1)。

• No growing KV cache / 无增长型 KV Cache — native RWKV inference updates a compact state instead of appending keys and values for every past token.
  原生 RWKV 推理持续更新紧凑状态，无需为每个历史 token 追加 Key 与 Value。

These properties make long-running streams and resource-constrained deployments easier to engineer. Read the original RWKV paper, the RWKV-7 paper, and the upstream RWKV-LM repository.

Complexity here is measured against sequence length. Model weights, activations, batch size, concurrent streams, and backend allocations still depend on the selected model and runtime.
这里的复杂度均以序列长度为参照；模型权重、激活、批量大小、并发流与后端分配仍取决于所选模型和运行时。

The architecture matters when it reaches real hardware. RWKV App brings model discovery, downloads, local inference, and multimodal workflows into one consistent product:

• Five operating systems — Android, iOS, Windows, macOS, and Linux
• Local model paths — after downloading supported weights, selected workflows can run on your device
• More than chat — explore text, vision, speech, translation, model comparison, and local AI workflows
• Hardware-aware acceleration — CPU, GPU, and selected NPU paths are available according to device, model package, and backend support

同一个开源客户端覆盖五个平台，并把模型管理、本地推理与多模态工作流统一在真实的端侧产品中。

flowchart LR
    A["RWKV models<br/>parallelizable training<br/>recurrent inference"] --> B["rwkv-mobile<br/>multi-backend runtime"]
    B --> C["rwkv_mobile_flutter<br/>Dart / FFI bridge"]
    C --> D["RWKV App<br/>five operating systems"]
    B --- E["CPU · WebGPU · QNN<br/>CoreML · MLX · MediaTek NPU"]

• RWKV-LM — the upstream architecture, research, training, and model ecosystem
• rwkv-mobile — the high-performance, multi-backend inference runtime
• rwkv_mobile_flutter — the Flutter adapter and Dart FFI bridge
• RWKV_APP — the cross-platform product, model management, and user experience

Backend availability varies by device and model package. See each runtime repository for the current support matrix.

1. Download RWKV Chat for your platform.
2. Choose a supported model that fits your device and download its weights.
3. Run locally and explore the workflows available for that model and backend.

• Android — Official download · Google Play · Latest release
• iPhone & iPad — App Store · TestFlight
• Windows, macOS & Linux — Official download · Latest release

• app_website — the RWKV Chat download website and supporting backend
• offline_reader — a browser extension that translates webpages through a local model server
• rwkv_studio — an offline-first workspace for models, providers, chat, and AI workflows
• rwkv_downloader — a focused download manager for model files

Explore all RWKV-APP repositories →

Try the app, report a reproducible issue, or improve one of the projects above.

• Report an issue or request a feature
• Join the Discord community
• 加入 QQ 技术交流群：325154699
• 加入 QQ 测试群：332381861

The core app is released under the Apache License 2.0. Other repositories may use different licenses; check each project before reuse.