Skip to content
Language models

Kaze-4.7B-A530M-MIT: A compact Sparse MoE language model for local devices

2 min readupdated

Kaze-4.7B-A530M-MIT is an experimental foundational language model trained from scratch by developer Kineyo under the MIT license. The model uses a Sparse Mixture-of-Experts architecture, meaning that out of a total size of about 4.72B parameters, only about 528M parameters are activated per token. It is optimized for local and private work on edge devices (CPU, Linux, and Android), and also comes in a ready-to-use GGUF format for llama.cpp and Kofu_Mobile. The model focuses on research into MoE architectures, as well as Spanish-language tasks and code generation.

Key features

  • Sparse MoE architecture: contains 1 shared expert and 31 routed experts (top_k = 2) with the SwiGLU activation function.
  • Hardware efficiency: around 528M active parameters per token lower computing resource requirements during inference.
  • Local execution: full support for running on CPU, Linux, and in the Kofu_Mobile mobile app thanks to pre-converted weights in GGUF format.
  • Context and tokenization: supports a context length of up to 4096 tokens and uses a custom Byte-Level BPE vocabulary of 32,768 tokens.

What you can do

  • CPU and terminal inference: run a quantized version of the model locally using llama.cpp for private text tasks.
  • Mobile app integration: use the model in Android projects through the Proyecto Kofu ecosystem.
  • Architecture research: study the training and operation features of compact Sparse MoE models built from scratch.

Example

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "shebxyd/Kaze-4.7B-A530M-MIT"
tok = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id, dtype=torch.bfloat16, device_map="auto")
prompt = "<|user|>\nHola, ¿quién eres y para qué sirves?\n<|assistant|>\n"
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=128, temperature=0.65, top_p=0.9, repetition_penalty=1.18, do_sample=True)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
Read moreDetails
Parameters: ~4.72B total / ~528M active
Layers: 23 layers, d_model = 1024, GQA 16Q / 4KV
Limitations: experimental model trained on an MIT-licensed corpus, with an emphasis on Spanish and code
File formats: BF16 safetensors and Q4_K_M GGUF quantization
Discuss Telegram 0 69

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 8.8 GB · ready files9.6 GB
  • FP8 / INT8recommendedweights 4.4 GB5.2 GB
  • GGUF Q4_K_Mweights 2.7 GB · ready files3.4 GB
  • 4-bitweights 2.5 GB3.3 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 3080

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

Language models: with a 8192-token context.

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

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent A30

    $0.26 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    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.

    # shebxyd/Kaze-4.7B-A530M-MIT — A30, FP8 / INT8
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve shebxyd/Kaze-4.7B-A530M-MIT --max-model-len 8192 --quantization fp8
    Add more LoRAs to Kaze-4.7B-A530M-MIT
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

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.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

New posts about versions of Kaze-4.7B-A530M-MIT (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

shebxyd/Kaze-4.7B-A530M-MIT
Author
shebxyd
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

690

Ask AI about this model

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

Comments