Alibaba's Qwen3.5 Beats GPT-5.2 on Key Benchmarks With 17B Active Parameters
Alibaba's new flagship Qwen3.5-397B-A17B unifies vision and text, activates just 17B of 397B parameters, and rivals GPT-5.2 and Claude 4.5 Opus at open-weight prices.
Takeaways
- Alibaba released Qwen3.5-397B-A17B, a 397B-parameter MoE with 17B active per token under Apache 2.0.
- First Qwen open model unifying vision, video, and text into one architecture, replacing separate Qwen3 and Qwen3-VL lines.
- Hybrid Gated DeltaNet plus 512-expert MoE architecture, up from 128 experts in earlier Qwen3 MoE releases.
- Scores 45 on Artificial Analysis Intelligence Index, ranking #3 among open-weights models per independent evaluation.
- Priced around $0.39 input and $2.34 output per million tokens on OpenRouter, with 262K native context extendable to 1M.
- Supports thinking and non-thinking modes in one checkpoint, plus tool calling via Qwen-Agent and MCP.
Alibaba's Qwen team just dropped a flagship that reshuffles the open-weights leaderboard. Qwen3.5-397B-A17B is a sparse Mixture-of-Experts model with 397 billion total parameters but only 17 billion activated per token, shipped under Apache 2.0. It is also the first Qwen model to natively handle images and video in the same architecture as text.
The model scores 45 on the Artificial Analysis Intelligence Index, ranking third among open-weights models behind GLM-5 and Kimi K2.5, while running with a fraction of the active parameters of its rivals. Kimi K2.5 uses 1T/32B, GLM-5 uses 744B/40B, and DeepSeek V3.2 uses 671B/37B, so Qwen3.5 is doing more with less compute per forward pass.
One model, two modalities, no more forks
The most consequential change is architectural consolidation. Qwen3.5 397B is the first Qwen open-weights release with native vision input, supporting image and video alongside text. Previously, Alibaba maintained separate model lines for vision (Qwen3-VL) and text-only (Qwen3), and this release folds those into a single checkpoint, following the broader industry move toward natively multimodal foundation models.
Reasoning and non-reasoning behavior are also collapsed into that same checkpoint. Where the Qwen3 family shipped as separate instruct and thinking variants, Qwen3.5 397B supports both modes internally. You toggle it at the API level rather than swapping weights.
What the hybrid architecture actually does
Under the hood, Qwen3.5 pairs a linear-attention block with a sparse MoE, an approach inherited from the experimental Qwen3-Next released last September. That earlier model was previewed but widely regarded as half-trained. Qwen3.5 takes the same architectural direction and scales it aggressively, jumping from 128 experts in the previous Qwen3 MoE models to 512 experts in the new release.