RWKV-7 G1l

RWKV-7 “Goose” is an attention-free recurrent language model. It maintains a constant-size recurrent state instead of a growing attention KV cache.

This release provides four G1l checkpoints dated 2026-10-31, converted from BlinkDL’s original BF16 PTH files. These are base language models, not safety-aligned assistants. The embedded chat template provides a conversational interface.

Available Tags

GGUF file size is not the same as total runtime RAM or VRAM use. Start with a size that fits your hardware.

Run

ollama run shoumenchougou/rwkv7-g1l:1.5b
ollama run shoumenchougou/rwkv7-g1l:2.9b
ollama run shoumenchougou/rwkv7-g1l:7.2b
ollama run shoumenchougou/rwkv7-g1l:13.3b

Choose one of the commands above. The included Modelfiles require Ollama 0.33.2 or later; this release was tested with Ollama 0.40.11.

Chat API

import requests

response = requests.post(
    "http://localhost:11434/api/chat",
    json={
        "model": "shoumenchougou/rwkv7-g1l:7.2b",
        "messages": [{"role": "user", "content": "What is deep learning?"}],
        "think": False,
        "stream": False,
    },
    timeout=300,
)
response.raise_for_status()
print(response.json()["message"]["content"])

Set think to True to request the embedded thinking prompt. Handling and display of reasoning text depend on the client and Ollama version.

Context Length

The published Modelfiles set num_ctx to 34304, matching the checkpoint training context.

Ollama may display 1M in the model list because GGUF architecture metadata declares a recurrent capacity field of 1,048,576. This is not a claim that these checkpoints were trained or tested at a one-million-token context. RWKV’s constant-size recurrent state does not by itself guarantee quality at arbitrary sequence lengths.

Source and Validation

The original checkpoints are pinned to BlinkDL/rwkv7-g1 revision 11e3c50128e479898eee79c48831a95cf3e21d91. Each source filename follows rwkv7-g1l-<size>-20261031-ctx34304.pth.

The GGUF files were quantized with llama.cpp b8419, commit 509a31d00f87804a466dee983a30ce29aa0441b1.

Validation on an NVIDIA RTX 4090 included native Transformers, the optimized Torch runtime, and FLA BF16 loading and generation, with finite-logit checks. The 13.3B BF16 tests used mixed GPU/CPU weight offload rather than claiming that the full BF16 model fits in 24 GB VRAM.

The selected quantizations passed Ollama GPU chat tests in thinking and non-thinking modes. Published registry manifests and blobs were downloaded into fresh verification directories and checked against their SHA-256 digests. The freshly downloaded GGUFs also passed separate llama.cpp CUDA generation tests. The fresh-download checks did not use a second Ollama daemon or a clean daemon model store.

These are functional smoke tests, not systematic quality benchmarks or a guarantee of parity across quantizations.

Other Formats

Each GGUF repository also provides FP16, Q4_K_M, Q5_K_M, Q6_K, and Q8_0 files. Model cards include source hashes, architecture details, and file-integrity information.

License and Limitations

Model weights are distributed under Apache License 2.0. RWKV names and logos remain subject to applicable brand and trademark rights.

The models may produce inaccurate, biased, unsafe, or otherwise undesirable text. Quantization can change their outputs. Do not use them as a substitute for qualified professional judgment in high-stakes settings.

RWKV website | RWKV-LM source | RWKV-7 paper