Instructions to use mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit with MLX:

# Make sure mlx-vlm is installed
# pip install --upgrade mlx-vlm

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

# Load the model
model, processor = load("mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit")
config = load_config("mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit")

# Prepare input
image = ["http://images.cocodataset.org/val2017/000000039769.jpg"]
prompt = "Describe this image."

# Apply chat template
formatted_prompt = apply_chat_template(
    processor, config, prompt, num_images=1
)

# Generate output
output = generate(model, processor, formatted_prompt, image)
print(output)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit"

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent

How to use mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit 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 "mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit"

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 mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit"

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 "mattbucci/Qwopus3.8-27B-Coder-Flash-MLX-4bit" \
  --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"

Qwopus3.8-27B-Coder-Flash — MLX 4-bit

MLX conversion of Jackrong/Qwopus3.8-27B-Coder-Flash, a coding-agent fine-tune of Qwen3.8-27B (dense Qwen3.5-class DeltaNet hybrid with the vision tower). 4-bit affine, group size 64, 15 GB on disk; runs in about 16 GB of unified memory plus KV.

Quantization recipe

Not a plain mlx_vlm.convert. The module classes that carry recurrent-state gate scalars and the vision tower stay in bf16, the rest is 4-bit:

kept bf16	quantized 4-bit g64
`linear_attn.in_proj_a` / `in_proj_b` (48 DeltaNet layers), all norms, `conv1d`, vision tower (27 blocks + merger + patch embed)	`lm_head`, `embed_tokens`, attention q/k/v/o (16 full-attention layers), `linear_attn.in_proj_qkv` / `in_proj_z` / `out_proj`, MLP gate/up/down (64 layers)

402 of 922 modules quantized, 4.7 bits per weight overall. Every quantized layer's weights, scales and biases pass a per-layer corruption scan. The processor_config.json that transformers 5.x writes (naming Qwen3VLImageProcessor, which AutoProcessor cannot resolve) is dropped; preprocessor_config.json resolves the Qwen2VL image processor.

Validation (Apple M4 Pro, 64 GB, SGLang + MLX)

gate	result
basic / tool call / thinking terminates	3/3
content-aware reasoning probe	VERIFIED
codegen probe (2 tasks, 8 tests)	8/8 STRONG
image probe	STRONG
video probe (both motion directions)	STRONG
decode tok/s @1K / 8K / 32K (single user)	14.6 / 14.2 / 13.2 (same as Qwen3.6-27B-4bit and Qwen3.8-27B-4bit on this stack)
SWE-bench Lite 3-instance sealed smoke (opencode)	3/3 non-empty patches on the canonical file (astropy 12907, 14182, 14365); two are the upstream change, the third the regex half of it
MMLU (100) / HumanEval (20) / LAB-Bench (50) / needle	83% / 90% / 37.5% / 4/4 at 1K, 4K, 16K, 64K (stock Qwen3.8-27B-4bit: 84% / 90% / 38.7% / 4/4)

The chat template always reasons: enable_thinking in chat_template_kwargs has no effect (same as the upstream Qwen3.8-27B template). Give coding requests a generous max_tokens; the model recommends sampling at temperature 1.0, top-p 0.95, top-k 20 (generation_config.json).

Use

pip install mlx-vlm
python -m mlx_vlm.generate --model <this repo> --prompt "Write a function that merges overlapping intervals." --max-tokens 2048

Served with SGLang's MLX backend (the stack this was validated on): MODEL=<this repo> scripts/launch.sh qwen36-27b from mattbucci/m4-sglang-inference, which carries the converter (scripts/quantize/convert_qwen36_inhouse.py) and the gate scripts.

Credits

Fine-tune by Jackrong on unsloth/Qwen3.8-27B (Qwen/Qwen3.8-27B); see the source card for the distillation recipe and its evaluations. Apache-2.0 throughout.