Instructions to use caslca/Qwen3.8-27B-mlx-uniform-4bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use caslca/Qwen3.8-27B-mlx-uniform-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("caslca/Qwen3.8-27B-mlx-uniform-4bit")
config = load_config("caslca/Qwen3.8-27B-mlx-uniform-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
Unsloth Desktop

How to use caslca/Qwen3.8-27B-mlx-uniform-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 "caslca/Qwen3.8-27B-mlx-uniform-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": "caslca/Qwen3.8-27B-mlx-uniform-4bit"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent

How to use caslca/Qwen3.8-27B-mlx-uniform-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 "caslca/Qwen3.8-27B-mlx-uniform-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 caslca/Qwen3.8-27B-mlx-uniform-4bit

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use caslca/Qwen3.8-27B-mlx-uniform-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 "caslca/Qwen3.8-27B-mlx-uniform-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 "caslca/Qwen3.8-27B-mlx-uniform-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"

Qwen3.8-27B-mlx-uniform-4bit

⚠️ Identity note (2026-08-26): the language trunk of this model is bit-identical to mlx-community/Qwen3.8-27B-4bit. A full-tensor md5 sweep over both repos shows 2179 of 2180 tensors identical; the single differing tensor is vision_tower.patch_embed.proj.weight, because this copy carries the restored bf16 vision tower (see below) while the mlx-community upload keeps the original. MLX uniform 4-bit (group size 64) quantization is deterministic, so two independent conversions of the same bf16 base produce the same weights. Text results measured against either repo apply to both; prefer this copy if you want the vision path.

27B parameters, vision tower included (restored 2026-08-23, see below) — note: Hugging Face's size badge undercounts packed 4-bit MLX weights (it counts the packed uint32 tensors), so the number shown beside this repo is wrong; the figure here is the true parameter count.

MLX uniform 4-bit quant of unsloth/Qwen3.8-27B.

measured	value
effective bits/weight	4.0 (uniform; all 498 quantized layers at 4-bit)
weights footprint	15.13 GB
quantized-layer bit histogram	4-bit: 498

Recommended sampling (measured, not vibes)

param	value
temperature	0.6 (certified by a per-model temperature ladder)
top_p / top_k / min_p	0.95 / 20 / 0.0
presence_penalty	0.0
max_tokens / thinking_budget	102400 / 81920 (thinking ON)

These values were certified by an execution-gated benchmark campaign (temperature ladders with convergence gates over HumanEval+/MBPP+ and agentic harnesses) — methodology and full results: https://github.com/ivan-avramov/mlx_local_stack.

Serving: MLX (mlx-lm / mlx-vlm). Quantized on-device with mlx_lm.convert (uniform) or mlx_optiq (mixed-precision KL-sensitivity recipes).

Native MTP drafter available (2026-08-26)

This checkpoint ships a native multi-token-prediction sidecar (optiq/mtp.safetensors, 29 tensors, outside the weight index — inert at normal load). A standalone, servable extraction is published at caslca/Qwen3.8-27B-mlx-uniform-4bit-mtp-drafter: measured 1.46–1.58× decode at 68–76% acceptance (probe-only — no quality certification; see that card's caveats).

Vision tower restored (2026-08-23)

The original conversion was language-model-only. This revision grafts the vision tower back from the upstream base (unsloth repackaging of the family release): 333 vision_tower.* tensors kept bf16 (exactly what the vision-retaining mlx_vlm convert produces for this family), +0.92 GB.

The text trunk is bit-identical to the evaluated artifact: the trunk shards are byte-copies (md5-verified), and a fixed-token forward pass through the language model produces bit-identical logits pre/post graft. Every benchmark number on this card measures exactly the weights this revision serves for text. One-image smoke passed post-graft.