RWKV-7 G1k

RWKV-7 “Goose” base language models in four sizes: 1.5B, 2.9B, 7.2B, and 13.3B.

RWKV-7 is an attention-free recurrent architecture. It maintains a constant-size recurrent state instead of an attention KV cache that grows with the sequence. G1k identifies this checkpoint generation from BlinkDL/rwkv7-g1.

Model lineup

Tag	Model size	Release quantization
`1.5b`	1.5B	Q6_K
`2.9b`	2.9B	Q4_K_M
`7.2b`	7.2B	Q4_K_M
`13.3b`	13.3B	Q4_K_M

Browse all four models on the Tags page.

The 1.5B release uses Q6_K. Larger releases use Q4_K_M, with token embeddings and output weights retained at Q6_K precision. Parameter labels identify the model size class; exact tensor counts and file checksums belong to each individual release.

Quick start

Install a current version of Ollama with RWKV-7 support. Local release checks use Ollama 0.34.4.

Run the 2.9B model:

ollama run shoumenchougou/rwkv7-g1k:2.9b

Or choose another size:

ollama run shoumenchougou/rwkv7-g1k:1.5b
ollama run shoumenchougou/rwkv7-g1k:7.2b
ollama run shoumenchougou/rwkv7-g1k:13.3b

Use an explicit size tag. The 1.5B model is the smallest option; larger models require more memory. Total runtime memory exceeds the downloaded model file size.

Prompt format and sampling

Each release uses the chat template embedded in its GGUF file, including the model’s thinking and non-thinking prompt forms. Chat formatting does not turn a base model into an instruction-tuned assistant.

Release sampling defaults:

Parameter	Value
Temperature	1.0
Top-k	128
Top-p	0.5
Frequency penalty	0.5
Presence penalty	0.5
Repeat penalty	1.0

The stop sequence is two newline characters (\n\n).

Context length

Context settings are specific to each checkpoint. Inspect the configured parameters of your selected model:

ollama show --parameters shoumenchougou/rwkv7-g1k:2.9b

Ollama may display 1M context because the GGUF architecture metadata records a capacity of 1,048,576. That figure is not the checkpoint’s training length or a claim of tested long-context quality. A constant-size recurrent state also does not imply unlimited recall.

Release validation

Each model is checked independently before publication: source SHA-256 and file size, tensor structure, quantization metadata, embedded template, Ollama parameters, model blobs, and short text generation.

Local checks run on CPU. They establish loading and generation compatibility; they do not establish full-context performance, CUDA benchmark results, or identical output across quantizations. Completed checks for one size should not be interpreted as results for another size.

Other formats

GGUF releases and checksums — FP16, Q4_K_M, Q5_K_M, Q6_K, and Q8_0.
Native Transformers safetensors.
FLA RWKV-7 collection.
ModelScope GGUF collection.

Some collections also retain earlier generations. Select entries explicitly named G1k.

Limitations and license

These are base language models, not safety-aligned assistants. They may produce inaccurate or undesirable text, and reliable instruction following is not guaranteed.

Model weights are distributed under Apache License 2.0. RWKV names and logos may be subject to separate trademark rules.

Models

Readme