Qwen3.5 Plus

Alibaba officially open-sourced Qwen3.5 on February 16, 2026, starting with Qwen3.5-397B-A17B, also named "Qwen3.5-Plus," as a natively multimodal foundation model. The model combines a linear attention mechanism with a sparse mixture-of-experts design to reduce compute requirements, delivering strong performance in re Qwen3.5-Plus supports 201 languages and dialects—up from 119 in the Qwen3 series—including additional low-resource languages such as Hawaiian, Fijian, and Niger-Congo languages. Across a broad set of benchmarks the model demonstrates strong performance in language understanding and reasoning, code generation, agentic w

Alibaba Group released Qwen3.5-Plus as an open-source foundation model on February 16, 2026, making it immediately available on its proprietary platforms and open-source communities including Hugging Face. The model matches the performance of Google's Gemini 3 Pro at a fraction of the cost, with Alibaba pricing the API Alibaba stated plans to roll out a more powerful flagship model, Qwen3.5-Max, along with other variants in the near future. The launch positions Qwen3.5-Plus as Alibaba's bid to deliver frontier-class multimodal performance at substantially lower inference cost, intensifying competition with Chinese rivals including By

Alibaba released Qwen3.5-Plus as an open-source foundation model on February 16, 2026, based on the Qwen3-Next architecture with a higher-sparsity Mixture-of-Experts design, Gated DeltaNet, and a hybrid attention mechanism combining Gated Attention. The model totals 397 billion parameters with 17 billion activated per The Baidu Baike entry also documents the early ecosystem adoption following launch: on February 17, 2026, Huawei's Ascend platform announced adaptation support for Qwen3.5-Plus enabling training and inference deployment, and on February 21, 2026, Alibaba Cloud officially launched the "Qwen Coding Plan" programming subs

The InferenceX technical profile confirms Qwen3.5-397B-A17B as the first open-weights release in Alibaba's Qwen3.5 series, announced by the Qwen team as a native vision-language foundation model with 397B total parameters and 17B activated per forward pass. The official announcement is dated February 2026 (Alibaba Clou Positioning is explicitly agentic and multimodal, targeting reasoning, coding, agent capabilities, and multimodal understanding, with strong generalization across agent tasks per OpenRouter's distribution description. The article reports early-fusion multimodal training achieves cross-generational parity with Qwen3 whi

AIBase's third-party news report covers Alibaba's release of Qwen3.5-Plus on February 16, 2026, describing it as a natively multimodal model with 397 billion total parameters and 17 billion activated parameters that is comparable to Gemini 3 Pro and has topped open-source large-model rankings. The article cites a 60% r The same report highlights Qwen3.5-Plus's multimodal generational leap via pre-training on mixed visual and text tokens, achieving top performance with less than 40% of the parameters of comparable models. It lists benchmark scores including 87.8 on MMLU-Pro (surpassing GPT-5.2), 88.4 on GPQA (outperforming Claude 4.5)

On February 16, 2026, Alibaba open-sourced Qwen3.5 as its latest model series, starting with the Qwen3.5-397B-A17B checkpoint also branded "Qwen3.5-Plus" as a natively multimodal foundation model with strong performance in reasoning, coding, agent capabilities, and multimodal understanding. The release combines a linea The announcement highlights Qwen3.5-397B-A17B's benchmark results across language understanding and reasoning, code generation, agentic workflows, image and video comprehension, and GUI interaction, positioning it as rivaling leading frontier models in both versatility and performance. The model's inference-efficiency

A Baidu Baike encyclopedia entry summarizes the Qwen3.5 family launch on February 16, 2026, covering two named versions, Qwen3.5-Plus and Qwen3.5-397B-A17B, which use a Hybrid Attention Mechanism and a Sparse Mixture of Experts (MoE) Architecture optimized for logical reasoning, mathematical computation, and code gener The entry reports an MMLU-Pro knowledge-reasoning score of 87.8 for Qwen3.5-Plus, claims an 8.6x inference-throughput increase at 32K context length, and lists Day-0 adaptation by NVIDIA, AMD, Apple, Huawei Ascend, Moore Threads, Muxi, and Haiguang. These claims are not independently corroborated by the supplied primar