Alibaba

Qwen3-Coder-Flash is a compact, 30B MoE model capable of local inference on modern MacBooks. It joins Alibaba’s broader Qwen3 ecosystem as a nimble counterpart to the heavyweight 480B hosted version, giving developers a pragmatic hybrid setup for coding workflows.

Alibaba has released two new open-source MoE models under its Qwen3 series, including an FP8 quantized variant. The 30B parameter architecture activates only 3B parameters per forward pass, enabling high performance with reduced hardware demand. With 128K context support and strong benchmarks in reasoning and code, these models are primed for local use cases on Apple Silicon Macs and beyond.

Text-to-video, image-to-video, and hybrid input modes now run locally at 720p and 24 fps. The Alibaba Wan team has open-sourced Wan 2.2 under Apache 2.0 and integrated it into ComfyUI and Hugging Face on launch day.

Qwen‑MT brings instruction-tuned translation across 92 languages, built on Qwen3. With daily model drops and Wan 2.2 on the horizon, Qwen is pushing a relentless pace in open AI development.

The Qwen team at Alibaba has launched Qwen3‑Coder, a massive open-source AI model designed for complex coding workflows. With 480 billion parameters and support for a 1 million token context window, the model challenges proprietary leaders like Claude Sonnet 4 while remaining freely available under Apache 2.0.

Alibaba Cloud has released a refined version of its flagship Qwen3‑235B model, specifically tuned for instruction-following tasks. The update separates the non-thinking mode, expands context length up to 1 million tokens, and delivers noticeable gains in reasoning, coding, and alignment benchmarks. The move simplifies deployment decisions for developers while solidifying Qwen’s standing in the open-source large language model arena.