Qwen3 Arrives and Local Inference Adapts

News

Alibaba Drops Qwen3 on Hugging Face: Thinking Model with 119 Languages

Qwen3 dropped this week, and the flagship model is a thinking model. This is a direct shot at DeepSeek R1's reasoning dominance, and it’s built on the Alibaba stack that already has strong local inference support.

MLX Deprecates OpenCL to Focus Strictly on Metal Backend

Apple is finally dumping the deprecated OpenCL backend and focusing strictly on Metal for MLX. Expect fewer abstraction bugs and faster performance on M-series chips moving forward.

Apple Hints at Open-Source LLM and VLM Weights for Developers

Apple is teasing an open-source LLM and a new VLM (Visual Language Model) framework. If they are actually going open-weight, this changes the game for on-device, privacy-first mobile apps.

Tools

vLLM Merges Potential Qwen3 MoE Support for Inference

The vLLM project is hardening support for Qwen3's mixture-of-expert architecture on NVIDIA and AMD hardware. If you are scaling up Qwen3 inference on a cluster, this is the backend patch you need to track.

Ollama v0.9.0 Ships with Native Qwen3 Thinking Support

Ollama v0.9.0 just dropped with full Qwen3 support, including the new thinking protocol. This is the fastest way to get the new models running locally without touching a Python environment.