Instructions to use eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX") prompt = "Write a story about Einstein" messages = [{"role": "user", "content": prompt}] prompt = tokenizer.apply_chat_template( messages, add_generation_prompt=True ) text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Pi
How to use eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX with Pi:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX"
Configure the model in Pi
# Install Pi: npm install -g @mariozechner/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "mlx-lm": { "baseUrl": "http://localhost:8080/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Hermes Agent new
How to use eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX with Hermes Agent:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX"
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:8080/v1 hermes config set model.default eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX
Run Hermes
hermes
- OpenClaw new
How to use eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX with OpenClaw:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX"
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:8080/v1 \ --custom-model-id "eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX" \ --custom-provider-id mlx-lm \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
- MLX LM
How to use eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Interactive chat REPL mlx_lm.chat --model "eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX"
Run an OpenAI-compatible server
# Install MLX LM uv tool install mlx-lm # Start the server mlx_lm.server --model "eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX" # Calling the OpenAI-compatible server with curl curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX", "messages": [ {"role": "user", "content": "Hello"} ] }'
Qwen3.6-35B-A3B K-Guard REAP 25% โ MLX 8-bit
English model card: README.md
์ด ๋ชจ๋ธ์
mlx-community/Qwen3.6-35B-A3B-8bit์
๊ธฐ๋ฐ์ผ๋ก ๋ง๋ ํ๊ตญ์ด ์ค์ฌ expert-pruned MLX ์ฒดํฌํฌ์ธํธ์
๋๋ค. Routed MoE
expert์๋ REAP ๋ฐฉ์์ router-weighted activation pruning์ ์ ์ฉํ์ผ๋ฉฐ,
๊ธฐ์กด 8-bit MLX ํ์, top-8 routing, shared expert, tokenizer, chat template๊ณผ
๋น-MoE ๊ตฌ์ฑ์์๋ ์ ์งํ์ต๋๋ค.
์ด ๋ฆด๋ฆฌ์ค์์ K-Guard๋ pruning ์ ์ routing ๋ฐ activation ํต๊ณ๋ฅผ ์์งํ ๋ ์ฌ์ฉํ ํ๊ตญ์ด ์ค์ฌ calibration profile์ ์๋ฏธํฉ๋๋ค. ๋ณ๋์ fine-tuning์ด๋ ๊ฐ์ค์น ์ฌํ์ต์ ํ์ง ์์์ต๋๋ค.
๋ชจ๋ธ ์์ฝ
| ํญ๋ชฉ | ๊ฐ |
|---|---|
| ๊ธฐ๋ฐ ์ฒดํฌํฌ์ธํธ | mlx-community/Qwen3.6-35B-A3B-8bit |
| ์ํคํ ์ฒ | Qwen3.6 MoE (qwen3_5_moe) |
| MoE layer | 40 |
| Layer๋น routed expert | 256๊ฐ์์ 192๊ฐ๋ก ์ถ์ |
| Layer๋น ์ ๊ฑฐ expert | 64 / 256 (25%) |
| Token๋น ํ์ฑ routed expert | 8๊ฐ, ์ ์ง |
| Shared expert | ์ ์ง |
| ์์ํ | MLX affine 8-bit, group size 64 |
| ๋ก์ปฌ ์ฒดํฌํฌ์ธํธ ํฌ๊ธฐ | ์ฝ 27.18 GiB |
| Pruning ํ ํ์ต | ์์ |
Calibration ๋ฐ expert ์ ์
์ด 24,576๊ฐ sequence๋ก ๊ตฌ์ฑ๋ ๊ฒฐ์ ์ ํ๊ตญ์ด ์ค์ฌ mixture์์ expert ์ค์๋๋ฅผ ๊ด์ธกํ์ต๋๋ค.
- ์ผ๋ฐ ํ๊ตญ์ด instruction ๋ฐ reasoning 8,192๊ฐ
- ํ์ ์๋ฏธ ๋์ pair 8,192๊ฐ
- ๋ฒ์ฉ capability replay 4,096๊ฐ
- Qwen3.6 text ๋ฐ reasoning ํนํ 4,096๊ฐ
Observer๋ seed 17, ์ต๋ sequence length 8,192๋ก 40๊ฐ MoE layer ์ ์ฒด๋ฅผ ๊ด์ธกํ์ต๋๋ค. ๊ฐ layer์์ router-weighted activation ํต๊ณ๋ก expert๋ฅผ ์ ๋ ฌํ์ฌ ์์ 192๊ฐ๋ฅผ ์ ์งํ๊ณ , router row๋ ๋์ผํ index๋ก ์ถ์ํ์ผ๋ฉฐ, top-8 routing์ ๊ทธ๋๋ก ์ ์งํ์ต๋๋ค.
Calibration ํ์ผ๊ณผ layer/expert๋ณ ์ ์ฒด observation CSV๋
eouya2/KGuard-Korean-MoE-Calibration-v1์
๊ณต๊ฐํ์ต๋๋ค.
ํ๊ฐ ๊ฒฐ๊ณผ
Base์ pruning ๋ชจ๋ธ ๋ชจ๋ ๋์ผํ deterministic subset๊ณผ prompt๋ก ํ๊ฐํ์ต๋๋ค.
| Benchmark | Base 8-bit | ์ด ๋ชจ๋ธ |
|---|---|---|
| KMMLU | 66.4 | 60.9 |
| MMLU-Pro | 64.3 | 58.8 |
| GSM8K | 96.7 | 96.7 |
| HumanEval+ pass@1 | 90.9 | 93.3 |
| ํ์ 10-benchmark macro | 81.2 | 79.7 |
KMMLU, MMLU-Pro, GSM8K๋ ๊ฐ๊ฐ 512๋ฌธํญ์ ๊ณ ์ subset์ ์ฌ์ฉํ๊ณ , HumanEval+๋ EvalPlus ์ ์ฒด 164๋ฌธํญ์ ์ฌ์ฉํ์ต๋๋ค. ์ ๊ฒฐ๊ณผ๋ ๋์ผํ ๋ก์ปฌ ํ๊ฐ ์กฐ๊ฑด์์ ๋ชจ๋ธ ๊ฐ ์ฐจ์ด๋ฅผ ๋น๊ตํ๊ธฐ ์ํ ์์น์ ๋๋ค.
MLX ์ฌ์ฉ๋ฒ
pip install -U mlx-lm
mlx_lm.generate \
--model eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX \
--prompt "ํ๊ตญ์ด๋ก mixture-of-experts ๋ชจ๋ธ์ ๊ฐ๋จํ ์ค๋ช
ํด ์ฃผ์ธ์." \
--max-tokens 256 \
--temperature 0.0
Python ์์ :
from mlx_lm import generate, load
model_id = "eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX"
model, tokenizer = load(model_id)
messages = [{"role": "user", "content": "ํ๊ตญ์ด๋ก ์๊ธฐ์๊ฐ๋ฅผ ํด ์ฃผ์ธ์."}]
prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=256))
OpenAI-compatible server:
mlx_lm.server --model eouya2/Qwen3.6-35B-A3B-KGuard-REAP25-8bit-MLX
๊ถ์ฅ ์ฉ๋
Apple Silicon ๋ก์ปฌ ์คํ, ํ๊ตญ์ดยทํ์ ์์ฑ, MoE pruning ์ฐ๊ตฌ, routed expert ์ฉ๋์ ๋ ์ ๊ทน์ ์ผ๋ก ์ค์ธ ๋ชจ๋ธ์ด ํ์ํ ์คํ์ ์ฌ์ฉํ ์ ์์ต๋๋ค. ์ค์ ์๋น์ค ์ ์ฉ ์ ์๋ ๊ธฐ๋ฐ ๋ชจ๋ธ๊ณผ ๋์ผํ๊ฒ ์ฉ๋๋ณ ํ์ง ๋ฐ ์์ ์ฑ ํ๊ฐ๋ฅผ ๊ถ์ฅํฉ๋๋ค.
์ฐธ๊ณ ์๋ฃ
- ๊ธฐ๋ฐ ๋ชจ๋ธ:
Qwen/Qwen3.6-35B-A3B - MLX 8-bit ๋ณํ:
mlx-community/Qwen3.6-35B-A3B-8bit - REAP ๋ ผ๋ฌธ: REAP the Experts
- MLX pruning ๊ตฌํ:
egesabanci/reap-mlx
๋ผ์ด์ ์ค
๋ชจ๋ธ์ upstream Qwen ๋ฆด๋ฆฌ์ค์ Apache 2.0 ๋ผ์ด์ ์ค๋ฅผ ๋ฐ๋ฆ ๋๋ค. ๋ณ๋ ๊ณต๊ฐํ ํฉ์ฑ calibration ๋ฐ์ดํฐ๋ CC BY 4.0์ ๋๋ค.