Qwen 3.6 & 3.5 Releases
Qwen announced Qwen3.6 Plus and Qwen3.5 Plus models, with the 3.5 release positioned as a cheaper option and both supporting images. The social post highlights that Qwen3.5 is about three times cheaper, supports images, and the models feature zero data retention; they are available in Go. (x.com)
Qwen has rolled out Qwen3.6-Plus and Qwen3.5-Plus, adding two hosted models that take text and images and target different price points. (qwen.ai) Qwen published Qwen3.5 on February 15, 2026, and launched Qwen3.6-Plus on April 1, 2026, with both offered through its application programming interface on Alibaba Cloud Model Studio. Qwen says each model comes with a 1 million token context window, which is the amount of text and other input a model can keep in working memory at once. (qwen.ai, qwen.ai) These are multimodal models, meaning one system can read words and inspect pictures in the same prompt instead of handing images off to a separate vision model. Qwen’s API platform lists Qwen3.5-Plus with text, image, and video inputs, and Alibaba Cloud’s model catalog says Qwen3.6-Plus and Qwen3.5-Plus support text, image, and video input. (qwen.ai, help.aliyun.com) Alibaba Cloud markets the split as a tradeoff between capability and cost: its model catalog labels Qwen3.6-Plus as balanced across quality, speed, and price, while the lower tier “Flash” line has already been upgraded to Qwen3.5. The same catalog says the commercial “Plus” endpoint has been upgraded to Qwen3.6 and the “Flash” endpoint to Qwen3.5. (help.aliyun.com) Qwen’s April 1 post frames Qwen3.6-Plus as a coding and agent upgrade, with stronger tool use, long-horizon planning, and multimodal reasoning. In its own benchmark table, Qwen reports Qwen3.6-Plus ahead of Qwen3.5-397B-A17B on SWE-bench Verified, Terminal-Bench 2.0, and several agent evaluations. (qwen.ai) The older Qwen3.5 release set the baseline for this push by introducing what Qwen called a native vision-language model, which means image understanding was built into the model rather than bolted on later. Qwen said that open-weight release, Qwen3.5-397B-A17B, expanded language and dialect support from 119 to 201. (qwen.ai) Pricing is published separately from the model launch posts, and Alibaba Cloud’s pricing page shows token-based billing that varies by deployment region and prompt length. The company’s English pricing page was updated April 1, 2026, and says Model Studio charges separately for input and output tokens, with batch and context-cache discounts on some models. (alibabacloud.com) On privacy, Alibaba Cloud says Model Studio “never trains on your data” and describes that policy as zero retention, alongside encryption and desensitization controls. That language matches the privacy claim circulating in social posts about the new Qwen hosted models. (alibabacloud.com) For developers, the practical point is that Qwen is pushing these models through OpenAI-compatible endpoints rather than a new interface. Alibaba Cloud’s February 10, 2026 software development kit page says developers can call Model Studio with OpenAI’s official software development kits in Python, Node.js, Java, and Go. (help.aliyun.com) That leaves Qwen with a simple pitch in April 2026: one newer “Plus” model for stronger agent and coding work, and one cheaper line for lower-cost multimodal calls, both plugged into the same application programming interface stack. (qwen.ai, help.aliyun.com)
