RWKV7-G1 GooseOne

RWKV7-G1 GooseOne

Pure RNN text-generation base models for chat, reasoning and tool calls, meant as a starting point for post-training and fine-tuning. Covers eight languages, for developers and researchers.

Made by BlinkDL on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: RWKV7-G1, nicknamed GooseOne, is a family of language models built as a pure RNN instead of a transformer. These are base models, pretrained on web, code and synthetic text plus instruction, chat and reasoning data. The author presents them as a starting point for post-training and fine-tuning.

Who it suits: It suits developers and researchers who want an RNN-based language model for chat, reasoning or tool calling, and who may fine-tune it themselves. Eight languages are listed: English, Chinese, French, Spanish, German, Portuguese, Russian and Italian.

Title: Who uses it
Industry | Who uses it | How they use it | Example
AI research | Model researchers | Fine-tune the base model or study its RNN design | Post-train it on their own chat data
Software | Application developers | Run it locally for chat or tool calling | Return a JSON function call from a tool list
Mobile and desktop apps | App builders | Run the GGUF builds in a local inference app | Chat with a model on a phone or laptop

People who want a ready-made assistant tuned for them can skip it, because these are base models.

Where to use it:

• Python with the rwkv package
• The Albatross project, for fast batch inference and a GUI
• The RWKV Chat app for mobile and desktop, found in the app stores
• Ollama, through the GGUF builds the card points to
• Windows, macOS or Linux, to download the files

Title: Top jobs
Use case | Who | What they get back
Chat | Developers | Text replies in a System, User and Assistant format
Reasoning | Researchers | An optional think step before the answer
Function calling | Agent builders | A JSON function call chosen from a tool list

How to set it up:

• Python on your computer
• The huggingface_hub package
• A runner for the weights, such as the rwkv package, Albatross or the RWKV Chat app

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Pick a runner from the card's list and point it at the files.
4. Strip any trailing space from your prompt, since a space at the end upsets the tokenizer.

pip install -U huggingface_hub
hf download BlinkDL/rwkv7-g1

How to use it: Write your prompt in the chat format below, ending on the Assistant line. The author notes that blank lines separate chat rounds in the training data, so replace blank lines inside your own text with single line breaks.

System: YOU_CAN_USE_SYSTEM_IF_NEEDED

User: PREVIOUS_STUFF

Assistant: PREVIOUS_STUFF

User: USER_PROMPT

Assistant:

For a hard question, the card suggests a think prompt. It also strongly recommends a fake think prompt, shown here.

User: USER_PROMPT

Assistant: <think></think

The card's decoding suggestions for the rwkv package come in two settings, one for chat and one for creative writing.

Chat: temp 1, topp 0.5, alpha_presence 2, alpha_frequency 0.1, alpha_decay 0.99

Creative (great for fiction etc.): temp 0.6, topp 0.6 ~ 0.8, alpha_presence 2, alpha_frequency 0.2, alpha_decay 0.99

Title: Projects the card names
Tool | What it does
Albatross | Efficient inference with a batch generation GUI
rwkv package | Python inference
RWKV Chat | Local inference app for mobile and desktop
RWKV-LM | Training code
RWKV-PEFT | Fine-tuning code

Example: For tool calling, you list three tools in the system line, covering weather, stock price and translation, and tell the model to return only a JSON function call. You then ask it to translate a short weather question into Japanese. It should answer with a JSON call to the translate tool. The card says this works with decoding set to zero for G1f at 1.5B and larger.

For newcomers and experts: If you are new, the easiest start is the RWKV Chat app, which runs a model locally on a phone or computer. A first try is a short chat message, with no trailing space after it.

For experts, this is an RNN, so the card says speed and memory use stay constant as the text grows. It lists decoding of a 7.2B model in fp16 on an RTX5090 at 145 tokens a second for one stream. The card also mentions DeepEmbed in the rwkv7a variant and tells you to prefer the latest models.

Download the weights
Pick a runner such as rwkv or Albatross
Send a prompt in the chat format
Read the reply or JSON call

Good to know:

• Base models: they are meant for post-training and fine-tuning, so expect rough edges as a chat assistant.
• Prompt format: a trailing space in your input can produce a non-English reply.
• License: Apache 2.0.
• Cost to run: none named. The card's speed figures come from an RTX5090 GPU.
• Data it reaches: it downloads files from Hugging Face, and the card names no other reach.
• Not for: tasks that need a ready-tuned assistant, since these are base weights.

Title: Decoding speed, 7.2B fp16 on RTX5090
One stream: 145 token/s
320 streams: 9650 token/s
960 streams: 10250 token/s

On Hugging Face, 13 November 2026

• Task: text generation
• Library: rwkv
• Downloads in the last 30 days: 9,911
• Likes: 253
• License: apache-2.0
• Languages: en, zh, fr, es, de, pt, ru, it
• Quantized versions: 40
• Fine-tunes: 42
• Tags: rwkv, pytorch, text-generation, causal-lm

Links

• Model: https://huggingface.co/BlinkDL/rwkv7-g1
• Files: https://huggingface.co/BlinkDL/rwkv7-g1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download BlinkDL/rwkv7-g1

Then follow the usage in its model card. Nothing in this kit runs by itself.