AI Radar: Qwen 3.6 — Alibaba's Open Middle Path

For years, one of the most active open-weight families has come out of China. With Qwen 3.6, though, the picture splits in two — and that split is the real story.

The 27B variant trails Qwen3.6 Plus by just three points — and is self-hostable as an open-weight model. (Quelle: Artificial Analysis — model leaderboard)

What's Open — and What Isn't

Qwen3.6-35B-A3B: a sparse mixture-of-experts model with 35B parameters, of which only about 3B are active per token. Strong at agentic coding and released under Apache 2.0 — so usable commercially.
Smaller variants: the 27B version is also openly available on Hugging Face and ModelScope.
Low active share: 3B active parameters means the model runs on comparatively modest hardware — the decisive point for SMEs.
The flagship stays closed: Qwen 3.7-Max and the Omni and Plus models are proprietary and reachable only through the Alibaba cloud.

Apache 2.0 — and Why It Counts

The license here is no footnote but the real story. Qwen 3.6 ships uniformly under Apache 2.0, and that is a deliberate break with its own past: with Qwen 2.5, Alibaba had placed the mid-size tiers — namely the 3B and 72B variants — under a more restrictive license that partly limited commercial use. For companies wanting to deploy a model productively and commercially, that was a genuine obstacle. With the 3.6 line, that hurdle falls away — concrete progress.

Apache 2.0 in practice means: commercial use explicitly allowed, no user thresholds, no pair-licensed contract, patent rights explicitly granted. Anyone running Qwen 3.6 on their own hardware may embed it in products, earn money from it, and pass it on — without legal gray zones. That is the same standard under which DeepSeek (MIT) or Mistral operate, and it lifts Qwen to a tier where you can reach for it without reservations.

The Caveat

Reaching for the closed flagship sends your inputs to an API in China — for personal or confidential data, a clear GDPR issue. The open 3.6 line sidesteps that, because it can run entirely in-house. That is exactly where its value lies for us.

Active vs. Total — the MoE Calculation

With mixture-of-experts, a closer look at the numbers pays off. The 35B parameters are the model's total size — they determine how much VRAM is needed simply to load the weights. The 3B active parameters, by contrast, set how much compute actually falls per token. That is the MoE basic equation: lots of capacity in the background, little consumption at the moment of the call. For self-hosting this means concretely: with four-bit or FP8 quantization, the 35B architecture fits on a single mid-range GPU, and inference stays economical enough for agentic runs across many steps.

Our Take

Qwen makes a direction visible: the strongest open model today is not the largest, but the most efficient. A model with 3B active parameters that handles agentic work cleanly and runs on your own hardware is, for many SME tasks, the more honest choice than a cloud giant.