Modelpublic

BlinkDL/rwkv7-g1

sourceHugging Faceapache-2.0updated 3d agoView on Hugging Face

228likes7.2kdownloads

Model Card

RWKV7-G1 "GooseOne" pure RNN reasoning model

These are BASE models (pretrained with web/code/synthetic + instruction/chat/reasoning data), suitable for post-training and fine-tuning (check https://huggingface.co/spaces/Jellyfish042/UncheatableEval to see their performance at language modeling).

More info & Gradio demo: https://rwkv.com/

Search "RWKV Chat" in play store / app store for our local inference app

RWKV Chat: https://rwkv.halowang.cloud/ (local inference for mobile/desktop) and https://github.com/RWKV-APP/RWKV_APP

Safetensor: https://huggingface.co/RWKV

GGUF: https://huggingface.co/collections/shoumenchougou/rwkv7-gxx-gguf

More GGUF & mobile models: https://huggingface.co/HaloWang/rwkv-weights (new) and https://huggingface.co/mollysama/rwkv-mobile-models (old)

Ollama GGUF: https://ollama.com/shoumenchougou (new) and https://ollama.com/mollysama (old)

RWKV-7 pth => GGUF script: https://github.com/MollySophia/rwkv-mobile/blob/master/converter/convertrwkvpthtogguf.py

Training: https://github.com/BlinkDL/RWKV-LM and https://github.com/Joluck/RWKV-PEFT

Note: rwkv7a has DeepEmbed

Efficient inference: https://github.com/BlinkDL/Albatross

—145+ token/s RWKV-7 7.2B fp16 bsz1 decoding @ RTX5090 (always const speed & vram)
—10250+ token/s RWKV-7 7.2B fp16 bsz960 decoding @ RTX5090 (always const speed & vram)
—9650+ token/s RWKV-7 7.2B fp16 bsz320 decoding @ RTX5090 (always const speed & vram)
—11289 token/s RWKV-7 7.2B fp16 bsz1 prefill @ RTX5090 (always const speed & vram)

pip inference: https://pypi.org/project/rwkv/

mobile inference: https://github.com/MollySophia/rwkv-mobile

There should not be any space at the end of your input (so strip your prompt) or you will upset the tokenizer and see non-English reponse.

Latest Best Practices for agentic work (study these projects): https://agent.objects.rwkvos.com/en

Eval example: https://github.com/BlinkDL/Albatross/blob/main/faster3a2605/evalgpqa_diamond.py

PROMPT GUIDE (including function call & agent): https://github.com/BlinkDL/RWKV-LM/blob/main/RWKV-v7/RWKV7-G1x-templates.txt

Function call: temp 0, topp 0, penalty 0, works for RWKV-7 G1f 1.5B and larger models:

System: Tools:
- get_weather(location: string, unit?: "celsius" | "fahrenheit")
- get_stock_price(ticker: string)
- translate_text(text: string, target_language: string)
Return only a JSON function call.

User: Translate "Will it rain tomorrow?" into Japanese.

Assistant: ```json

and

System: Tools:
[
{"name":"find_free_slots","description":"Find free calendar slots","arguments":{"date":{"type":"string"},"duration_minutes":{"type":"integer"},"time_window":{"type":"string"}}},
{"name":"create_calendar_event","description":"Create a calendar event","arguments":{"title":{"type":"string"},"start_time":{"type":"string"},"end_time":{"type":"string"},"attendees":{"type":"array","items":{"type":"string"}}}}
]
Return only a JSON function call.

User: Schedule a 30-minute sync with Bob on 2026-05-08 afternoon.

Assistant: ```json
{"name":"find_free_slots","arguments":{"date":"2026-05-08","duration_minutes":30,"time_window":"afternoon"}}

User: Function output: {"freeslots":[{"start":"2026-05-08T15:00:00+09:00","end":"2026-05-08T15:30:00+09:00"}],"bobemail":"bob@example.com"}

Assistant: ```json

The key is to keep it concise. Here you can enable "\<think\>" for Assistant too.

You can also use this template, which is closer to training data:

Assistant: <think> ... </think> <toolcall> {...} </toolcall> <toolcall> {...} </toolcall>

User: <toolresponse> {...} </toolresponse> <toolresponse> {...} </toolresponse>


---

Please always use **latest models** because they are always better at everything.

Decoding Suggestion (note: this is for RWKV pip pkg, which apply temp after topp):

Chat: temp 1, topp 0.5, alphapresence 2, alphafrequency 0.1, alpha_decay 0.99

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


Chat prompt (note: better replace all \n\n in USER_PROMPT to \n as i am using \n\n as "chat round separator" in pretrain data):

System: YOUCANUSESYSTEMIF_NEEDED

User: PREVIOUS_STUFF

Assistant: PREVIOUS_STUFF

User: USER_PROMPT

Assistant:


Think prompt (for hard prompts):

User: USER_PROMPT

Assistant: <think

**Fake think prompt (great result, highly recommended)**:

User: USER_PROMPT

Assistant: <think></think

---
Think prompt, alternative style, **for G1c and newer** models. Note there is a space before the "(think)" after USER_PROMPT:

User: USER_PROMPT (think)

Assistant: <think

Shorter think, same style:

User: USER_PROMPT (think a bit)

Assistant: <think

Longer think, same style:

User: USER_PROMPT (think a lot)

Assistant: <think

---

FIM prompt (**for G1c and newer** models, works for text & code & everything):

Better (recommended):

✿prefix✿✿suffix✿and that’s how first I got interested in AI research.✿middle✿When I was young, I only liked to```


Note "✿" will always be tokenized to one single token in RWKV tokenizer, so I picked it.

---

0.1B = L12-D768 0.4B = L24-D1024 1.5B = L24-D2048 2.9B = L32-D2560 7.2B = L32-D4096 13.3B = L61-D4096 Vocab = 65536 for all current models Head size = 64 for all current models

Gxx = Data Version

G0x = less than 1 epoch, as training 1 epoch for a large model is expensive :( G0 G0a G0a2 G0a3 ... G0b ... = adding more (newer and better) data, so G0a has better quality (but less) data than G1

G1x = more than 1 epoch G1 G1a G1a2 G1a3 ... G1b ... = adding more (newer and better) data, note G1a has better quality (and more) data than G0a