Skip to content
Other

noesis-rwkv7-g1d-0.4b-ticks

an experiment with internal ticks in RWKV

2 min readupdated

A demonstration checkpoint of the experimental noesis method, which gives a recurrent model time inside its own state and trains it to use that time through distillation. It is based on the lightweight rwkv7-g1d-0.4b base model (0.4 billion parameters) from RWKV authors (BlinkDL). The model stands out by running the vector of the final layer back into the input state multiple times ("ticks") without token generation before producing an answer, which helps prevent repetition loops and yields more coherent English and Russian text. This development will benefit language model architecture researchers and anyone experimenting with recurrent models.

Key features

  • Bilingual support: Handles everyday questions and generates coherent answers in both English and Russian.
  • Improved stability: Unlike the base model, it avoids infinite repetition loops ("Green joy... green trees") thanks to internal ticks.
  • Lexical diversity: Demonstrates an increased number of unique words per response compared to the base version (48.5 versus 42.3).
  • Efficient latent space: The effective rank of the latent representation increases during ticks, preventing collapse to a single point.

What you can do

  • Test chat probe: Run the chat probe script from the noesis repository to check model responses with varying tick counts.
  • Study architecture: Explore stateful embedding code and the practical mechanics of RWKV-7 recurrent states.
  • Compare behavior: Evaluate the generation difference between the base model and the 4-tick checkpoint on your own prompts.

Example

Sample response to the question "What is your favorite color and why?":
Base model (no ticks): "I love green. It symbolizes: ... Green joy (green flowers) ..." (loops)
This checkpoint (4 ticks): "Green is the color I love. It symbolizes purity, joy, transparency, and a bright future..."

Read moreDetails
Benchmarks: on test task families, ticks do not yet add fundamentally new logical power (44 solved tasks out of 304 versus 46 for the base).
Limitations: without ticks, the checkpoint wraps responses in the service tag , so the model is designed specifically to work with ticks.
Training details: 100 steps of full fine-tuning (Muon) across 4 ticks with distillation of responses from the tickless model (KL at temperature 2) without an external teacher.
Discuss Telegram 0 27

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 1.7 GB · ready files2.3 GB
  • FP8 / INT8recommendedweights 0.8 GB1.4 GB
  • 4-bitweights 0.5 GB1.0 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # Vaniello/noesis-rwkv7-g1d-0.4b-ticks
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download Vaniello/noesis-rwkv7-g1d-0.4b-ticks --local-dir ~/models/noesis-rwkv7-g1d-0.4b-ticks
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of rwkv7-g1 (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face

About the model

Vaniello/noesis-rwkv7-g1d-0.4b-ticks
Author
Vaniello
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

270

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments