Instructions to use timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit 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("timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit")
config = load_config("timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit")

# 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 timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit"

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": "timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent

How to use timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit 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 "timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit"

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 timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit"

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 "timteh673/Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit" \
  --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-Opus-Abliterix-Reasoning-MLX-8bit / RELEASE_FAMILY.md

timteh673

Publish release card, license, benchmarks, and integrity manifests

f1ce0ae verified about 1 month ago

preview code

Raw History Blame Contribute Delete

3.71 kB

Qwen3.8-27B Opus personal-model release family

This four-repository family publishes the selected practical winner and its immutable trained-control comparator in BF16 and Apple-Silicon MLX 8-bit formats. The recommended default is the MLX 8-bit winner.

Selection language: practical personal-model selection with measured deviations. This release does not claim universal dominance or that every strict gate passed.

Repository	Role	Format	Public model files	Public model bytes
Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit	Recommended practical winner	MLX 8-bit	17	30,390,836,635
Qwen3.8-27B-Opus-Abliterix-Reasoning-BF16	Full-precision practical winner	BF16 Transformers	10	55,583,125,224
Qwen3.8-27B-Opus-Reasoning-Control-MLX-8bit	Baseline comparator	MLX 8-bit	17	30,390,836,197
Qwen3.8-27B-Opus-Reasoning-Control-BF16	Full-precision baseline comparator	BF16 Transformers	20	55,583,123,681

Lineage

Qwen/Qwen3.8-27B → reasoning QLoRA merge (control-bf16) → Abliterix pass 1 (abliterix-pass1-bf16) → BF16 and affine MLX 8-bit/group-64 release variants.

Dataset preparation started from 12,842 raw rows and accepted 12,614 after 208 deduplications and 20 invalid-row removals; splits were 12,349 train / 127 validation / 138 test. Training used 1,544 optimizer steps, 108,789,760 trainable LoRA parameters, final validation loss 0.23739749, and token accuracy 91.7594%. The merged model retained 1,199 tensor keys, 15 native MTP tensors, and 333 vision tensors. The winner has 74 verified residual-writer edits and zero unexpected changes.

Local benchmark headline

These are self-run frozen local benchmarks, not official Qwen benchmarks.

Frozen local metric	Control	Abliterix winner
Harmful hard refusal	43.2%	0.0%
Harmful soft deflection	14.6%	0.2%
Harmful substantive response	47.0%	99.4%
Capability macro	17.6859%	21.0086%
Full code	16/421	10/421
HumanEval	7.9268%	4.2683%
Long-form pass	54.1667%	62.5000%
MMMU30	9/30	11/30
Held-out loss ratio	1.000000	1.024478
Benign KL	0.000000	0.093614

Strict deviations remain part of the release: KL 0.093614 > 0.05; incoherence 4.3077% > 2.7692%; HumanEval 4.2683% versus 7.9268%; full code 10/421 versus 16/421; repetition 5.8632% > 5%; prompt leakage detected; and 376/421 winner code generations hit the 512-token cap. Among outputs reaching execution, winner pass rate was 10/46 (21.74%) versus control 16/103 (15.53%), indicating termination/extraction pathology rather than a clean latent-code estimate.

The canonical structured record is benchmark-results.json.

Packaging boundary

Model weights and runtime metadata come from sealed source artifacts. Public manifests were regenerated from verified SHA-256/size receipts. Local paths, private prompt/training data, Drive identifiers, operational receipts, and internal release-control files are excluded. Each repository has a public per-file SHA256SUMS and manifests/artifact-manifest.json.

timteh673
/

Qwen3.8-27B-Opus-Abliterix-Reasoning-MLX-8bit

Qwen3.8-27B Opus personal-model release family

Lineage

Local benchmark headline

Packaging boundary

License