RWKV-7 G1i

RWKV-7 G1i is a family of RWKV-7 “Goose” language models packaged for easy local inference with Ollama.

RWKV-7 is an attention-free recurrent architecture with a constant-size recurrent state. Unlike Transformer-based models, it does not require a KV cache that grows with context length, while training remains parallelizable.

Available models

Quick start

Run the 2.9B model:

ollama run shoumenchougou/rwkv-7-g1i:2.9b

Or use the Ollama API:

curl http://localhost:11434/api/chat \
  -d '{
    "model": "shoumenchougou/rwkv-7-g1i:2.9b",
    "messages": [
      {
        "role": "user",
        "content": "Explain why RWKV uses constant state."
      }
    ]
  }'

Python:

from ollama import chat

response = chat(
    model="shoumenchougou/rwkv-7-g1i:2.9b",
    messages=[
        {
            "role": "user",
            "content": "Explain why RWKV uses constant state."
        }
    ],
)

print(response.message.content)

About RWKV-7

RWKV-7 “Goose” combines recurrent inference with highly parallelizable training.

Key properties include:

• Attention-free recurrent architecture
• Constant-size recurrent state
• Constant recurrent memory requirements during generation
• No growing Transformer-style KV cache
• Designed for efficient local and long-context inference

The G1i checkpoints contain web, code, synthetic, instruction, chat, and reasoning training data.

Model information

This Ollama repository provides quantized versions of the RWKV-7 G1i model family.

For the original checkpoints, architecture details, prompt guidance, training information, and full model documentation, see:

• RWKV organization on Hugging Face
• RWKV-7 G1 model repository
• RWKV-7 “Goose” with Expressive Dynamic State Evolution
• RWKV-LM on GitHub

Notes

These models are provided for local inference and experimentation. Model quality, reasoning behavior, instruction following, and language performance vary by model size, quantization, and prompt.

For tasks requiring higher capability, prefer the larger 7.2B or 13.3B variants when hardware resources allow.