@Lonely__MH: Unleashed! The uncensored version of Qwen3.8-27B with safety restrictions removed is here! Kudos to the community for the speed! Deeply optimized for Mac M chips! I see everyone discussing the DGX Spark deployment for ling-3.0-flash, and many people's first reaction is that the compute power is too expensive to buy. Since cloud costs are high...

X AI KOLs Timeline Models
uncensored-ai mac-optimization local-deployment multimodal-model safety-research qwen3 mlx

Summary

Qwen3.8-27B uncensored version released, optimized for Mac M chips, supports local deployment, retains multimodal capabilities and safety research features, with simplified installation steps.

Do as you wish! The uncensored version of Qwen3.8-27B with safety restrictions removed is here! Kudos to the community for the speed! Deeply optimized for Mac M chips! I see everyone discussing the DGX Spark deployment for ling-3.0-flash, and many people's first reaction is that the compute power is too expensive to buy. Since cloud costs are high, why not use your Mac to locally deploy this true local multimodal powerhouse: Qwen3.8-27B-Uncensored-MLX Key points: 1. 0 rejection: Orthogonal removal of rejection directions, directly complies with various extreme red team/safety research instructions. 2. All capabilities retained: Multimodal image understanding, thinking control, tool calling, and 262K long context are all present. 3. High-quality weights: Core visual tower maintains BF16 precision, only language linear layers are Affine quantized. 4. Storage requirements: Recommended 4-bit version (only 15GB space, runs smoothly on Mac with 24G memory); lossless 8-bit optional (27.5GB space, requires 32G memory) - everyone choose according to their capacity. 3-step minimal local deployment (compatible with OpenAI interface): (1) Upgrade library: pip install -U mlx-vlm (2) Download 4-bit weights -- hf download orcarouter/Qwen3.8-27B-Uncensored-MLX --include "4-bit/*" --local-dir ./Qwen3.8-27B-Uncensored-MLX -- (3) One-click start local API server -- python -m mlx-vlm server --model ./Qwen3.8-27B-Uncensored-MLX/4-bit --port 8080 -- Local deployment, safe and uncensored, truly do as you wish!
Original Article
View Cached Full Text

Cached at: 08/18/26, 06:30 PM

Fully Unleashed! Introducing the Uncensored Version of Qwen3.8-27B with Safety Restrictions Stripped Away!

Hats off to the community’s speed! Deeply optimized for Mac M-series chips!

Many are discussing the deployment of ling-3.0-flash on DGX Spark, but the first reaction is often that the computing power is too expensive to afford.

Since cloud costs are high, why not deploy this true local multimodal powerhouse directly on your Mac: Qwen3.8-27B-Uncensored-MLX.

Key highlights:

  1. Zero Refusal: Orthogonally decoupled from the refusal direction, directly complying with various extreme red-team/safety research instructions.

  2. Full Capability Retention: Multimodal image understanding, thought control, tool use, and a massive 262K long context are all preserved.

  3. High-Quality Weights: The core vision tower maintains BF16 precision, with only the language linear layers undergoing Affine quantization.

  4. Memory Requirements: Recommended 4-bit version (requires only 15GB of disk space, runs smoothly on 24GB RAM Macs); for lossless quality, the 8-bit option is available (27.5GB disk space, requires 32GB RAM) - choose based on your setup.

3-Step Simple Local Deployment (OpenAI API Compatible):

(1) Upgrade the library:
pip install -U mlx-vlm

(2) Download the 4-bit weights:

hf download orcarouter/Qwen3.8-27B-Uncensored-MLX --include "4-bit/*" --local-dir ./Qwen3.8-27B-Uncensored-MLX

(3) Launch the local API server with one command:

python -m mlx-vlm server --model ./Qwen3.8-27B-Uncensored-MLX/4-bit --port 8080

Local deployment, completely uncensored. It’s truly unleashing the full potential.

Similar Articles

@nicekate8888: For the past twenty days, I've been obsessing over one thing — how to make Qwen3.6-27B run fast and well on my Mac. I started with Unsloth Q5, got 18 tok/s, and the fan was roaring. Then I switched to MLX 6bit + DFlash, hitting 22 tok/s, still not fast enough. Eventually I found MTPLX 4bit: 43 tok/s with good quality.

X AI KOLs Timeline

The user shares their experience optimizing Qwen3.6-27B inference speed on a Mac using different quantization methods (Unsloth Q5, MLX 6bit + DFlash, MTPLX 4bit), ultimately reaching 43 tok/s.

@Lonely__MH: Come on! Do as you please 2.0 version is here! Qwen 3.8-27B cracked version direct API call! No local deployment needed! Completely free your computer! Yesterday, I shared a local tutorial for Qwen-3.8 27B uncensored version, and many followers found it too high of a barrier. So, I'm bringing the API right to you…

X AI KOLs Timeline

This article announces the cracked version of Qwen 3.8-27B, offering direct API call functionality without local deployment, aiming to simplify access to AI models.

@Xudong07452910: A hot comment section on Hacker News: Qwen 3.6 27B is the ideal choice for local development. Key findings: dense parameter model, native support for 256k context, running Q8_0 quantized version at 30 tokens/…

X AI KOLs Timeline

Qwen 3.6 27B is a dense 27B model that achieves impressive performance on local hardware with 256k context, running at 30 tokens/s on MacBook Max M5 and 50 tokens/s on RTX 5090, and is considered by some as the first local model with true general intelligence.