PHMay bagong interesting model sa Hugging Face: Qwen3.8-27B-Ùn*énsøréd-MLX from OrcaRouter.
Basically, ito ay an abliterated/refusal-removed version ng Qwen3.8-27B, converted to MLX para mas practical i-run locally sa Apple Silicon Macs.
Ang interesting dito:
Pero important: Ùn*énsøréd doesn't mean better.
The safety alignment has been substantially removed, meaning the model can generate content that the original Qwen would refuse. The creators specifically position it for AI-safety research, red-teaming, interpretability, and controlled experiments.
For developers with a Mac, mas interesting yung MLX implementation dahil you can experiment with a relatively large multimodal model locally without relying entirely on a cloud API.
Basically:
Local LLM + Vision + Reasoning + Tool Calling + No built-in refusal layer.
Interesting model to experiment with, but definitely one where you need your own safety and moderation layer if you're building anything on top of it.
Check it out:
Basically, ito ay an abliterated/refusal-removed version ng Qwen3.8-27B, converted to MLX para mas practical i-run locally sa Apple Silicon Macs.
Ang interesting dito:
- 27B parameters
- Native vision-language capabilities
- Reasoning at tool-calling
- Supports 2-bit, 4-bit, 6-bit, and 8-bit quantization
- 4-bit version is around 15 GB
- Designed for Apple Silicon / MLX
- Puwedeng gamitin with MLX-VLM, LM Studio, Pi, Hermes Agent, at OpenClaw
- 4/6/8-bit versions reportedly passed the uploader's refusal probes with zero refusals
Pero important: Ùn*énsøréd doesn't mean better.
The safety alignment has been substantially removed, meaning the model can generate content that the original Qwen would refuse. The creators specifically position it for AI-safety research, red-teaming, interpretability, and controlled experiments.
For developers with a Mac, mas interesting yung MLX implementation dahil you can experiment with a relatively large multimodal model locally without relying entirely on a cloud API.
Basically:
Local LLM + Vision + Reasoning + Tool Calling + No built-in refusal layer.
Interesting model to experiment with, but definitely one where you need your own safety and moderation layer if you're building anything on top of it.
Check it out: