Alibaba releases the Qwen3-VL vision models, the Qwen3Guard βsafety moderationβ models, and three closed-weight models, including Qwen3-Max with 1T+ parameters
Qwen 50.6kΒ βΒ Safetensors qwen3_vl_moe Julian Nabil / Forbes Middle East : Alibaba Introduces Qwen3-Max AI Model With Over 1T Parameters Markus Kasanmascheff / WinBuzzer : Alibaba Releases Qwen3-VL Open-Source Vision Language AI Model Series Ankush Das / Analytics India Magazine : Alibaba Launches Qwen3-VL With Open Source Flagship Model Reuters : Alibaba launches Qwen3-Max AI model with more than a trillion parameters X: @alibaba_qwen : We're excited to announce the upgrade of Qwen3-Coder, and the upgraded API βqwen3-coder-plusβ is now available on Alibaba Cloud Model Studio with major improvements: π» Enhanced terminal task capabilities and better performance on Terminal Bench (w/ Qwen Code / Claude Code) π [image] Ethan Mollick / @emollick : So far, Qwen3-Max seems impressive for a non-reasoning model, doing a good job at a lot of my weird tests that even some reasoners struggle with. [image] @alibaba_qwen : π Introducing Qwen3-LiveTranslate-Flash β RealβTime Multimodal Interpretation β See It, Hear It, Speak ItοΌ π Wide language coverage β Understands 18 languages & 6 dialects, speaks 10 languages. ποΈ VisionβEnhanced Comprehension β Reads lips, gestures, onβscreen text and [image] @alibaba_qwen : π‘οΈ Meet Qwen3Guard β the Qwen3-based safety moderation model series built for global, real-time AI safety! π Supports 119 languages and dialects β 3 sizes available: 0.6B, 4B, 8B β‘ Low-latency, Real-time streaming detection with Qwen3Guard-Stream π Robust Full-context safety [image] Bindu Reddy / @bindureddy : The leaders in open source are far and away Qwen and DeepSeek US still lags way behind in this category For example - we still have to all the way back to Llama, if we need a fine-tune based on a US base model Awni Hannun / @awnihannun : Just for fun, here's what 32 simultaneous long-context generations with Qwen3 Next 80B looks like on an M3 Ultra. Using the new batch generation in mlx-lm. Context size for each is about 5k tokens: [video] @teortaxestex : This is what innovation looks like in ML. Lots of small things combined. Qwen3 VL is SOTA. And yet... always something subtle missing with these guys. It's the first model that has seen *a* connection and dismissed it in favor of the cat's psyop. But its vision is clear at least. [image] Omar Khattab / @lateinteraction : Sort of like calling a Qwen3-235B-A22B βqwen 50Bβ for short. Ahmad / @theahmadosman : qwen3 omni technical paper summary > qwen3-omni is one model for everything > text, vision, audio, speech, and video > beats chatgpt 4o and gemini 2.5 in reasoning and recognition > 30B model, only 3B active parameters per token > runs on consumer hardware with ease, usable [image] Andi Marafioti / @andimarafioti : Qwen3-Omni is here, and it's a huge step for omni-modal AI. π₯ Give it anything and get back text or surprisingly natural speech. Just watch the demo: it reads data from a table and then analyzes a snowboarder's balance in a video. The future is now. π€― [video] Junyang Lin / @justinlin610 : This is the 1st shot! For a long time people just don't have any idea about the safety work that we have invested efforts in. This time, we show you our safety guard model,Qwen3 Guard, specifically including generative guard Qwen3Guard-Gen and streaming guard model with @_akhaliq : qwen3-coder-plus is now available on Anycoder Enhanced terminal task capabilities and better performance on Terminal Bench (w/ Qwen Code / Claude Code) SWE-Bench performance up to 69.6 Safer code generation available as Qwen3-Coder-Plus-2025-09-23 [video] Elvis / @omarsar0 : Qwen3-Omni Technical Report A unified multimodal model that matches same-size Qwen text-only and vision-only baselines while pushing audio and audio-visual SOTA. Key technical details below: [image] Tianbao Xie / @tianbaox : After another half year, we are glad to bring Qwen3-VL! It's definitely the best open model you can access to start your digital agent and physical agent journey. Thanks the whole team! @shuai_bai_ @huybery @DunjieLu1219 @xuhaiya2483846 @JustinLin610 Junyang Lin / @justinlin610 : This is the 5th shot! Super crazy! We opensourced a 235B-A22B Instruct and Thinking Qwen3-VL models under Apache 2.0! Qwen3-VL, the new generation of our vision-language model, whose previous version was released a long time ago. During these days, we have conducted a lot of @sixsigmacapital : $BABA This could be Alibaba's mini chat-GPT moment. Aran Komatsuzaki / @arankomatsuzaki : RLPT: Reinforcement Learning on Pre-Training Data β’ RL directly on pre-train data (no human labels) β’ Next-segment reasoning objective (ASR + MSR tasks) β self-supervised rewards β’ Gains on Qwen3-4B: +3.0 MMLU, +8.1 GPQA-Diamond, +6.6 AIME24, +5.3 AIME25 [image] AshutoshShrivastava / @ai_for_success : In the last 12 hours, Qwen has released: > Qwen3Guard > Personal AI Travel Designer > Qwen3-LiveTranslate-Flash > Upgrade Qwen3-Coder > Qwen3-VL-235B-A22B > Qwen3-Max The Qwen team is crazy π₯ [image] @tryagentsea : Alibaba has: - best open weights image model (qwen image) - best open weights image editing model (qwen image edit 2509) - best sota open weights vision model (qwen3 vl) - best open weights video inpainting model (wan 2.2 animate) - one of the best foundation models (qwen3 max) Justine Chang / @justine_chang39 : Ok just did my image cropping test on Qwen3 VL It is, ON PAR, if not BETTER than Gemini 2.5 Pro for my use case π€― This is the FIRST non-Gemini model to be able to do this. This is really really good!! @Alibaba_Qwen @huybery @JustinLin610 [image] @reach_vb : NEW: Qwen 235B A22B Vision Language Model is OUTT! Apache 2.0 licensed and upto 1 Million context length π€― https://huggingface.co/... Merve / @mervenoyann : my vibe tests with Qwen3-Omni family of models > document performance with Instruct is very good π― > video understanding is nice β―οΈ > Thinking performs better in English > I suggest to use Captioner if you really want audio output, other two hallucinates a bit [video] @alibaba_qwen : π Qwen3-Max is hereβno preview, just power! Qwen Chat: https://chat.qwen.ai/ Blog: https://qwen.ai/... API: https://www.alibabacloud.com/ ... We've supercharged coding & agentic skillsβnow Qwen3-Max-Instruct without thinking rivaling top models on SWE-Bench, Tau2-Bench, [image] Chujie Zheng / @chujiezheng : Qwen3-VL, this is what you many guys are always wanting. Enjoy π» @alibaba_qwen : π We're thrilled to unveil Qwen3-VL β the most powerful vision-language model in the Qwen series yet! π₯ The flagship model Qwen3-VL-235B-A22B is now open-sourced and available in both Instruct and Thinking versions: β Instruct outperforms Gemini 2.5 Pro on key vision [image] Binyuan Hui / @huybery : We have released Qwen3-Max, the most powerful Qwen model to date! By continuously scaling up model size, data, and RL tasks, great things have happened. This time, coding and agent capabilities have also been significantly enhancedβenjoy! Jianwei Yang / @jw2yang4ai : πExcited to see Qwen3-VL released as the new SOTA open-source vision-language model! What makes it extra special is that it's powered by DeepStack, a technique I co-developed with Lingchen, who is now a core contributor of Qwen3-VL. When Lingchen and I developed this technique Bluesky: Sung Kim / @sungkim : Chinese AI has caught up with leading U.S.-based labs.Β Alibaba's release of Qwen3-Max places it alongside frontier AI players like Anthropic, Google, OpenAI, and xAI.Β βΒ qwen.ai/blog?id=2413...Β [images] Tim Kellogg / @timkellogg.me : Qwen3-Max: Just Scale ItΒ βΒ it's now safe to say Qwen is a frontier labΒ βΒ qwen.ai/blog?id=2413...Β [image] Threads: NaveedUllah / @naveed_ullah600 : Qwen has unveiled π€ππ²π»π― π§π§π¦, a next gen text to speech model built for natural, expressive audio.Β It supports English + multiple Chinese dialects (Mandarin, Cantonese, Beijing, Shanghai, Sichuan) and delivers voices that sound truly human-like. β¦ Mastodon: Simon Willison / @simon@fedi.simonwillison.net : Plus notes on the 5 (!) new things Qwen released today, the most exciting of which is the first in their Qwen3-VL vision-LLM series, a 235B 471 GB Apache 2 licensed monster! https://simonwillison.net/... Forums: Hacker News : Qwen3-VL
Context & Ripple Effects
Alibaba has been extending Qwen across visual reasoning, agentic coding and multimodal inputs: Qwen2.5-VL added device-control and vision tasks, while Qwen3-Coder targeted agentic software work. This release turns those adjacent capabilities into a broader product lineup.
The mix of an Apache-licensed vision flagship, safety models and closed-weight offerings clarifies a two-track approach: wide model access alongside proprietary cloud-oriented products.
First-order effects
- Developers can deploy or adapt the open-weight Qwen3-VL flagship, while teams needing moderation gain Qwen3Guard options for batch and streaming use across supported languages.
- Alibaba expands its commercial model catalog with closed-weight Qwen3-Max and an upgraded Qwen3-Coder API on Alibaba Cloud Model Studio, concentrating its most proprietary coding and agent features in its service layer.
Second-order effects
- Model buyers gain more leverage to compare open vision-language systems against proprietary alternatives, while vendors competing for enterprise workloads must match not only model capability but safety tooling and deployment choice.
- Bundling coding, vision, translation and moderation around Alibaba Cloud Model Studio can make the cloud platform a more natural destination for customers that start with Qwen models, reinforcing distribution beyond the base model itself.
Third-order effects
- If this split between open foundation models and closed premium services persists, competition will increasingly center on the surrounding stackβsafety, APIs, tools and hostingβrather than on weight availability alone.
- The addition of real-time moderation and interpretation points toward multimodal AI being sold as operational infrastructure, where reliability and governance become purchase criteria alongside benchmark performance.
The trend: AI labs are industrializing model portfolios by pairing open releases that broaden adoption with proprietary services that capture higher-value production workloads.