From their post:
Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.
Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation.
- Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables.
- Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery.
More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀


Same here. The small MoE ones are the only reason local is useful for me at all. I run qwen3:30b-a3b through Ollama on a MacBook and get roughly 70–95 tok/s, which is fast enough that I hand it all the boring jobs: summarising logs, pulling fields out of emails, first-pass triage. Anything that needs judgement still goes to a bigger model, because the a3b will happily do a tidy job of the wrong thing if the instruction is even slightly ambiguous. So I spot-check everything it does.
The other half of my setup is an old i7 home server with no usable GPU, and the same model is close to useless there. So what I’m really waiting for from Qwen4 is a small model that’s smarter per active parameter, not a bigger one. If they ship a ~30B-A3B-class Qwen4 that follows instructions better, that’s an instant swap for me. At least until I can spring for the Strix Halo!