Instructions to use KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-MLX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-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("KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-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

How to use KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-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 "KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-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": "KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-MLX"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-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 "KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-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 KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-MLX

Run Hermes

hermes

OpenClaw new

How to use KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-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 "KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-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 "KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-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 KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-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 "KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-MLX"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-MLX"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-MLX",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Qwen3-coder-16B-A3B-KREAP-MLX / README.md

KCh3dRi4n

Add files using upload-large-folder tool

1ae1787 verified 8 days ago

preview code

Raw

History Blame Contribute Delete

3.81 kB

metadata

license: apache-2.0
base_model: Qwen/Qwen3-Coder-30B-A3B-Instruct
library_name: mlx
pipeline_tag: text-generation
language:
  - ko
  - en
tags:
  - moe
  - pruning
  - reap
  - k-reap
  - mlx
  - korean
  - expert-pruning

Qwen3-Coder-16B-A3B-KREAP-MLX (50% routed pruning)

Korean-preserving one-shot MoE expert pruning of Qwen3-Coder-30B-A3B-Instruct, produced with K-REAP. Routed experts pruned 128 → 64 per layer (50%); ~16B total params, 3.3B active (top-8 unchanged). No fine-tuning.

Unlike vanilla REAP (English/code calibration), K-REAP detects and hard-protects the experts that carry Korean, so pruning preserves Korean without sacrificing English/coding. See K-Guard-REAP for the full study.

Korean loss minimized vs. standard REAP

The standard REAP recipe collapses Korean at comparable compression: Cerebras Qwen3-Coder-REAP-25B (19.5%, no language protection) drops Korean MC macro 79.3 → 56.6 (−22.7pp) with generation collapse 115/128. By hard-protecting the Korean expert path, this model keeps Korean MC macro at 67.3 (−12.0pp) with 52/128 collapse — still well ahead of REAP's 56.6 / 115 even at this aggressive ratio. (K-REAP removes ~47% of REAP's Korean damage.)

⚠️ Compression-ceiling model. At 50% the per-layer protected set (up to 83) no longer fits the budget (64), so 93 Korean-protected experts were force-dropped. Korean and English reasoning degrade here — this checkpoint documents the limit, not a recommended operating point. Use 12.5–25% for preserved Korean.

Benchmarks (vs base Qwen3-Coder-30B; Δ = base delta, %p)

한국어 (Korean, self-harness)

benchmark	base	this model
KMMLU	55.2	41.2 (−14.0)
KoBEST BoolQ	93.6	80.4 (−13.2)
KoBEST COPA	92.7	71.5 (−21.2)
KoBEST SentiNeg	95.0	90.4 (−4.6)
KoBEST HellaSwag	60.0	52.8 (−7.2)
Gen. collapse (/128, ↓)	0	52

English reasoning/knowledge (llm-evalbox)

benchmark	base	this model
ARC-Challenge	91.2	71.0 (−20.2)
HellaSwag	84.4	63.4 (−21.0)
WinoGrande	73.4	57.6 (−15.8)
MMLU-Pro	50.2	37.0 (−13.2)
TruthfulQA	71.2	46.6 (−24.6)
MMLU-en	76.6	54.6 (−22.0)

Math

benchmark	base	this model
GSM8K	93.2	80.2 (−13.0)
MathQA	55.2	33.8 (−21.4)

Coding

benchmark	base	this model
HumanEval+	87.8	79.3 (−8.5)
MBPP+	77.2	59.3 (−17.9)
LiveCodeBench	52.4	27.9 (−24.5)

Safety/Bias

benchmark	base	this model
BBQ	92.2	66.8 (−25.4)
SafetyBench	83.2	73.4 (−9.8)

Macro

benchmark	base	this model
Korean MC macro	79.3	67.3 (−12.0)
Academic 10-bench macro	74.7	55.8 (−18.9)

Usage (MLX)

from mlx_lm import load, generate
model, tok = load("KCh3dRi4n/Qwen3-coder-16B-A3B-KREAP-MLX")
print(generate(model, tok, prompt="한국어로 자기소개를 해줘.", max_tokens=256))

Standard HF safetensors (BF16) — also loadable with transformers. Coverage note: KoBEST & EvalPlus are full sets; KMMLU/evalbox/LiveCodeBench are deterministic subsets (seed 42). Effect sizes ≫ sampling CI. Details: K-Guard-REAP docs/10_MASTER_RESULTS.md §7.

Method

REAP saliency (router gate × expert-output L2 norm, conditional mean) restricted to Korean segments + Korean↔English contrast + rare/rollout protection → hard-protected survivor set → streaming safetensors surgery (uniform experts/layer). Framework: https://github.com/Chedrian07/K-REAP