Model Watch · Reviewed

Alibaba Qwen3.5

Announced Feb 16, 2026Released Feb 16, 2026Reviewed Jun 23, 2026
What they claimed

Alibaba's Qwen team positioned Qwen3.5 as a flagship open-weight refresh, headlined by Qwen3.5-397B-A17B — a sparse Mixture-of-Experts model with ~397B total parameters but only ~17B activated per token — built on a hybrid attention design (Gated DeltaNet plus sparse MoE) for large efficiency gains. The company claimed the flagship beats its own much larger trillion-parameter Qwen3-Max while running far cheaper, and that it outperformed GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro on a majority of vision benchmarks. Qwen3.5 was pitched as natively multimodal and agent-oriented, supporting 201 languages and a 262K context window extensible past 1M tokens. It launched as a full family spanning sub-1B to 397B sizes.

What shipped

Alibaba released Qwen3.5 as genuinely open-weight under Apache 2.0, distributing models via Hugging Face, ModelScope, and GitHub, with the smaller sizes following shortly after the flagship. API pricing was aggressive at roughly $0.40/$2.40 per million input/output tokens internationally, undercutting comparable Western frontier APIs by a wide margin.

The verdict

Qwen3.5 is the clearest sign that the open-weight frontier has a serious, well-resourced Chinese leader, and unlike many hype cycles the weights, license, and full size range actually shipped. The sparse-MoE efficiency story is the substance worth keeping: activating ~17B of 397B parameters delivers strong results at a fraction of the cost and serving footprint of dense trillion-parameter models, and the small models punching above their size class is a real capability for edge and on-device use. The vendor's 'beats GPT-5.2 / Claude / Gemini' framing should be read with the usual skepticism — those wins are concentrated on selected vision benchmarks against rivals with undisclosed and likely larger parameter counts, not a blanket superiority claim. Apache 2.0 licensing makes it genuinely usable and self-hostable, which is the durable differentiator. As with other Chinese open models, the strategic caveat for executives is sourcing and governance, not capability or license terms.

Why it matters

Qwen3.5 gives enterprises a permissively licensed, self-hostable model family efficient enough to run from phones to data centers, making open-weight AI a credible alternative to paid frontier APIs for many workloads.

Sources
  1. Qwen — Qwen3.5 official blog
  2. DeepLearning.AI The Batch — Qwen3.5 open-weights MoE family