Qwen3.5 Plus

Alibaba Unveils Qwen3.5-Plus, Undercutting Gemini 3 Pro on Cost - Chinese tech giant says new open-source AI model rivals Google’s flagship system while charging a fraction of the API price

Baidu Baike documents Qwen3.5-Plus as an open-source large language model released by Alibaba on February 16, 2026, belonging to the new generation of the Qianwen (Qwen) series. The model is based on the Qwen3-Next architecture and incorporates a higher-sparsity MoE design plus a hybrid attention mechanism combining Ga As a native multimodal large language model, Qwen3.5-Plus achieves deep integration of vision, language, and code through early text-visual fusion technology, with pre-training based on visual and text mixed tokens. Across multiple benchmarks, its performance is comparable to closed-source models such as GPT-5.2 and Ge

DataCamp's technical explainer provides the clearest model/service distinction: Qwen3.5-397B-A17B is the open-weight model released under Apache 2.0, while Qwen3.5-Plus is a hosted, API-only proprietary service built on the same underlying model and accessed via Alibaba Cloud Model Studio on pay-per-token pricing. The The hosted Qwen3.5-Plus service extends the context window to 1 million tokens, targeting long-context applications. The model is positioned as a general-purpose foundation model for both consumer and enterprise use cases, designed for native multimodal and agentic workflows. DataCamp frames Qwen3.5 as Alibaba's respon

InferenceX provides a technical breakdown of Qwen3.5-397B-A17B, the first open-weights release in Alibaba's Qwen3.5 series, published on Hugging Face as a post-trained checkpoint compatible with Transformers, vLLM, SGLang, and KTransformers under the Apache 2.0 license. A hosted counterpart called Qwen3.5-Plus offers a Architecturally, the checkpoint unifies thinking and non-thinking behavior in a single model: it operates in thinking mode by default, emitting content before the final answer, and unlike Qwen3 it drops the and soft switches, requiring API parameters to obtain non-thinking responses. The model uses early-fusion vision-

OpenRouter lists Qwen3.5 Plus 2026-02-15 as a native vision-language model with a hybrid linear-attention + sparse MoE architecture, supporting a 1M context window and offering in/out pricing of $0.26 / $1.56 per 1M tokens, hosted solely by Alibaba Cloud International with 100.00% uptime. Performance metrics show a P50 The page also reports an Auto-routing benchmark score of 44.8% and an average tool-call error rate of 6.37% for the Alibaba Cloud Int. provider, while throughput percentiles (P50 52 tok/s, P75 58 tok/s) and latency percentiles (P50 1.18s, P99 17.70s) give developers concrete operational signals for integration planning