Alibaba’s Qwen lineup has evolved rapidly over the past few weeks. We recently saw Qwen3-Coder-Next targeting developers with an AI coding assistant. This was followed by Qwen Image 2.0, which pushed the platform’s image generation quality even further. Each release strengthened a specific capability within the ecosystem. Now, building on that evolution, comes the Qwen 3.5 family with two new AI models – its first open weight model: the Qwen3.5 397B-A17B, and the Qwen3.5-Plus.
Among the two, the former, or the Qwen3.5 397B-A17B, is the flagship model, while the Qwen3.5-Plus is the hosted model available via Alibaba Cloud Model Studio. Both models can now be accessed on the Qwen Chat.
From what Alibaba tells us, the Qwen 3.5 family focuses on stronger reasoning, coding, agentic capabilities, multimodal understanding, and improved efficiency. More importantly, it reflects a broader push by Alibaba toward AI systems that can handle complex, multi-step tasks with greater autonomy. If you look at it carefully, the model is more than just an upgrade – it is a signal of where the Qwen family is heading.
In this article, we cover what’s new in Qwen 3.5, where it stands competitively, and what our hands-on testing reveals about its real-world performance. Let’s jump right in.
Table of contents
What is Qwen 3.5?
Qwen 3.5 isn’t just “the next Qwen model.” Alibaba has officially kicked off the Qwen 3.5 series by open-sourcing the first model, and has officially named it ‘Qwen3.5-397B-A17B.’
Now here’s the most important part, as far as its functioning goes – the model has 397 billion total parameters, but it doesn’t use all of them every time. Thanks to a sparse Mixture-of-Experts (MoE) setup, it activates only 17B parameters per forward pass. This is a fancy way of saying: big brain, but it only “wakes up” the parts it needs, so inference stays fast and cost-efficient.
Even more importantly, this is a native vision-language model. This means that it is built to handle text + images together, not as an afterthought. Alibaba claims it performs strongly across reasoning, coding, agent capabilities, and multimodal understanding in benchmark evaluations.
And there’s a very “real-world” upgrade too: language support jumps from 119 to 201 languages and dialects, which matters if you’re building any global-facing apps.
In parallel, Alibaba has also announced Qwen3.5-Plus, which is a hosted version available via Alibaba Cloud Model Studio. It offers a 1 million-token context window by default and includes built-in tools with adaptive tool use. This makes it suitable for long-context workflows and agent-style automation.
This brings us to the question – how does Qwen 3.5 do all this? Let’s have a look under its hood to understand this.
Under the Hood: How Qwen 3.5 Works
Qwen 3.5 is interesting not just because of its size, but how efficiently it uses that scale.
At the infrastructure level, the model separates how vision and language components are processed instead of forcing them into a one-size-fits-all pipeline. This heterogeneous setup allows text, images, and video inputs to be processed more efficiently, enabling near-100% training throughput even on mixed multimodal data.
Efficiency is further boosted by sparse activations. This allows different components to compute in parallel. Add to that a native FP8 pipeline – applying low precision where safe while preserving higher precision in sensitive layers – and the system cuts activation memory by roughly 50% while improving speed.
Alibaba also built a scalable asynchronous reinforcement learning framework to continuously refine the model. By separating training and inference workloads, the system improves hardware utilization, balances load dynamically, and recovers quickly from failures. Techniques like speculative decoding, rollout replay, and multi-turn rollout locking further improve throughput and stability, especially for agent-style workflows.
Pretraining: Power, Efficiency, and Versatility
Qwen 3.5 was pretrained with a clear focus on three things: power, efficiency, and versatility.
It was trained on a significantly larger mix of visual and text data than Qwen 3, with stronger multilingual, STEM, and reasoning coverage. Despite activating only 17B parameters at a time, the model reportedly matches the performance of much larger trillion-parameter systems.
Architecturally, it builds on the Qwen3-Next design, combining higher-sparsity MoE with hybrid attention mechanisms. This allows dramatically faster decoding speeds while maintaining comparable performance.
The model is also natively multimodal, fusing text and vision early in training. Language coverage expands from 119 to 201 languages and dialects, while a larger 250k vocabulary improves encoding and decoding efficiency across languages.
Benchmark Performance: Where Qwen 3.5 Stands
Benchmarks show us where a model begins to separate itself from the herd of options out there. Based on Alibaba’s released evaluations, Qwen3.5-397B-A17B delivers competitive performance across reasoning, agentic workflows, coding, and multimodal understanding. Here is a look at its benchmarks and what it means:
Instruction Following & Reasoning
- IFBench (Instruction Following): 76.5 — among the top scores in its class
- GPQA Diamond (Graduate-level reasoning): 88.4 — competitive with frontier reasoning models
These results suggest strong comprehension and structured reasoning which are critical for real-world workflows.
Agentic & Tool Use Capabilities
- BFCL v4 (Agentic tool use): 72.9
- BrowseComp (Agentic search): 78.6
- Terminal-Bench 2 (Agentic terminal coding): 52.5
Qwen 3.5 performs especially well in agent-driven tasks, reinforcing its positioning for workflow automation and tool orchestration.
Coding & Developer Workflows
- SWE-bench Verified: 76.4
This places it solidly in the range of models capable of handling real coding and debugging workflows.
Multilingual Knowledge
- MMLU: 88.5
The score aligns with its expanded language coverage and improved knowledge retrieval.
Multimodal & Visual Reasoning
- MMMU-Pro (Visual reasoning): 79.0
- OmniDocBench v1.5 (Document understanding): 90.8
- Video-MME (Video reasoning): 87.5
- VITA-Bench (agentic multimodal interaction): 49.7
These numbers highlight one of Qwen 3.5’s biggest strengths: multimodal comprehension across documents, visuals, and video.
Embodied & Spatial Reasoning
- ERQA: 67.5
This reflects improving capabilities in real-world and embodied reasoning scenarios.
What These Benchmarks Really Mean
Instead of dominating a single category, Qwen 3.5 shows balanced strength across reasoning, agentic execution, coding, and multimodal understanding. That balance matters because modern AI workloads aren’t single-task problems. They involve tools, documents, images, code, and multi-step workflows, and Qwen 3.5 appears to be built for exactly that reality.
Hands-on With Qwen 3.5
We conducted a couple of tests on both Qwen3.5 397B-A17B and the Qwen3.5-Plus. Here are the tests and the results.
Task 1 – Coding with Qwen3.5-Plus
Prompt:
You are an expert frontend developer and UI/UX designer.
Build a modern, responsive promotional website (single-page landing site) for the following event. The site should be visually premium, conversion-focused, and optimized for registrations.
Event Details:
Title: iqigai AI Fellowship Challenge 2026
Tagline: India’s Largest AI and Data Tech Hunt
Presented by: Fractal
Partner: Analytics Vidhya
Registration Link:
https://analyticsvidhya.com/datahack/contest/iqigai-genai-fellowship-challenge/?utm_source=social&utm_medium=X&utm_campaign=postContent to Include:
– Headline: India’s Largest AI and Data Tech Hunt is now live!
– Description:
The iqigai AI Fellowship Challenge 2026 is more than a hackathon — it’s a career-defining platform where participants compete, get nationally ranked, and gain visibility among top employers.
– Dates: 20th January – 8th March 2026
– Total Prize Pool: ₹20 Lacs
– Top Prizes:
Winner – ₹5 Lakhs
1st Runner-up – ₹3 Lakhs
2nd Runner-up – ₹2 LakhsWebsite Requirements:
1. Use HTML, CSS, and JavaScript (or React if preferred).
2. Fully responsive (desktop + mobile).
3. Modern gradient/AI-tech themed styling.
4. Smooth scrolling navigation.
5. Clear CTA buttons linking to registration page.
6. Sections:
– Hero section (large headline + CTA)
– About the Challenge
– Key Highlights / Why Participate
– Prize Section (cards or visual badges)
– Timeline / Dates
– Call-to-Action Banner
– FooterDesign Guidelines:
– Dark tech gradient background
– Subtle animations / hover effects
– Clean typography
– Cards with shadows and rounded corners
– Optional icons or illustrations
– Maintain professional event branding toneOutput Requirements:
– Provide complete runnable code
– Organize clearly into files
– Comment important parts
– Do NOT include placeholder lorem ipsum
– Ensure production-ready structureGenerate the full website code now.
Output:
Task 2 – Text-to-image with Qwen3.5-Plus
Prompt:
Create a cinematic anime-style transformation scene featuring Vegeta from Dragon Ball Super unlocking Ultra Ego — depict a dark cosmic battlefield as his body radiates destructive god-like ki, muscles tightening and posture shifting into fierce confidence, hair turning deep purple and eyes glowing magenta, surrounded by a raging flame-like violet aura that crackles and distorts the environment; capture the essence of a God-of-Destruction mindset where power grows through battle intensity and damage, emphasizing savage pride, chaotic energy waves, shattered terrain, and dramatic lighting — ultra-detailed, high contrast, dynamic camera angles, motion blur, and explosive anime shading, conveying overwhelming destructive dominance and unstoppable escalation.
Output:
Task 3 – Image-to-video with Qwen3.5-Plus
Simply click the Create Video option on the Image
Output:
Task 4 – Text-to-image with Qwen3.5 Open Weight
Prompt:
“Slash and Burn” could be a spirit or force of nature, embodying the cycle of destruction and renewal. It might appear as a fiery, elemental being that consumes everything in its path, only for new life to emerge from the ashes. This entity could be worshipped or feared as a deity of transformation and rebirth. bottom left signature “sapope”
Output:
Task 5 – Image-to-video with Qwen3.5 Open Weight
Simply click the Create Video option on the Image
Output:
Final Video:
Conclusion
The Qwen 3.5 family, with Qwen3.5 Open, is a step toward a more capable, unified AI system. With its hybrid MoE architecture, native multimodal design, expanded language coverage, and strong performance across reasoning, coding, and document understanding benchmarks, Alibaba is clearly optimizing for real-world workloads.
What stands out most is the balance. Instead of excelling in one narrow task, Qwen 3.5 shows consistent strength across agentic workflows, multimodal reasoning, and efficiency at scale. As AI moves from chat interfaces to execution-driven systems, models built for versatility and throughput will matter more. With the benchmark performances and the results we see in our hands-on tests, Qwen 3.5 positions itself firmly in that future.
