Alibaba’s Qwen lineup has advanced quickly over the previous few weeks. We not too long ago noticed Qwen3-Coder-Subsequent focusing on builders with an AI coding assistant. This was adopted by Qwen Picture 2.0, which pushed the platform’s picture technology high quality even additional. Every launch strengthened a particular functionality inside the ecosystem. Now, constructing on that evolution, comes the Qwen 3.5 household with two new AI fashions – its first open weight mannequin: the Qwen3.5 397B-A17B, and the Qwen3.5-Plus.
Among the many two, the previous, or the Qwen3.5 397B-A17B, is the flagship mannequin, whereas the Qwen3.5-Plus is the hosted mannequin out there by way of Alibaba Cloud Mannequin Studio. Each fashions can now be accessed on the Qwen Chat.
From what Alibaba tells us, the Qwen 3.5 household focuses on stronger reasoning, coding, agentic capabilities, multimodal understanding, and improved effectivity. Extra importantly, it displays a broader push by Alibaba towards AI techniques that may deal with complicated, multi-step duties with higher autonomy. When you take a look at it rigorously, the mannequin is extra than simply an improve – it’s a sign of the place the Qwen household is heading.
On this article, we cowl what’s new in Qwen 3.5, the place it stands competitively, and what our hands-on testing reveals about its real-world efficiency. Let’s soar proper in.
What’s Qwen 3.5?
Qwen 3.5 isn’t simply “the following Qwen mannequin.” Alibaba has formally kicked off the Qwen 3.5 collection by open-sourcing the primary mannequin, and has formally named it ‘Qwen3.5-397B-A17B.’
Now right here’s an important half, so far as its functioning goes – the mannequin has 397 billion whole parameters, nevertheless it doesn’t use all of them each time. Due to a sparse Combination-of-Specialists (MoE) setup, it prompts solely 17B parameters per ahead move. This can be a fancy means of claiming: massive mind, nevertheless it solely “wakes up” the elements it wants, so inference stays quick and cost-efficient.
Much more importantly, it is a native vision-language mannequin. Which means that it’s constructed to deal with textual content + photos collectively, not as an afterthought. Alibaba claims it performs strongly throughout reasoning, coding, agent capabilities, and multimodal understanding in benchmark evaluations.
And there’s a really “real-world” improve too: language assist jumps from 119 to 201 languages and dialects, which issues should you’re constructing any global-facing apps.
In parallel, Alibaba has additionally introduced Qwen3.5-Plus, which is a hosted model out there by way of Alibaba Cloud Mannequin Studio. It gives a 1 million-token context window by default and contains built-in instruments with adaptive instrument use. This makes it appropriate for long-context workflows and agent-style automation.
This brings us to the query – how does Qwen 3.5 do all this? Let’s take a look beneath its hood to know this.
Below the Hood: How Qwen 3.5 Works
Qwen 3.5 is fascinating not simply due to its dimension, however how effectively it makes use of that scale.
On the infrastructure degree, the mannequin separates how imaginative and prescient and language elements are processed as a substitute of forcing them right into a one-size-fits-all pipeline. This heterogeneous setup permits textual content, photos, and video inputs to be processed extra effectively, enabling near-100% coaching throughput even on combined multimodal knowledge.
Effectivity is additional boosted by sparse activations. This enables totally different elements to compute in parallel. Add to {that a} native FP8 pipeline – making use of low precision the place protected whereas preserving increased precision in delicate layers – and the system cuts activation reminiscence by roughly 50% whereas bettering pace.
Alibaba additionally constructed a scalable asynchronous reinforcement studying framework to constantly refine the mannequin. By separating coaching and inference workloads, the system improves {hardware} utilization, balances load dynamically, and recovers rapidly from failures. Strategies like speculative decoding, rollout replay, and multi-turn rollout locking additional enhance throughput and stability, particularly for agent-style workflows.
Pretraining: Energy, Effectivity, and Versatility
Qwen 3.5 was pretrained with a transparent concentrate on three issues: energy, effectivity, and flexibility.
It was skilled on a considerably bigger mixture of visible and textual content knowledge than Qwen 3, with stronger multilingual, STEM, and reasoning protection. Regardless of activating solely 17B parameters at a time, the mannequin reportedly matches the efficiency of a lot bigger trillion-parameter techniques.
Architecturally, it builds on the Qwen3-Subsequent design, combining higher-sparsity MoE with hybrid consideration mechanisms. This enables dramatically sooner decoding speeds whereas sustaining comparable efficiency.
The mannequin can also be natively multimodal, fusing textual content and imaginative and prescient early in coaching. Language protection expands from 119 to 201 languages and dialects, whereas a bigger 250k vocabulary improves encoding and decoding effectivity throughout languages.
Benchmark Efficiency: The place Qwen 3.5 Stands
Benchmarks present us the place a mannequin begins to separate itself from the herd of choices on the market. Primarily based on Alibaba’s launched evaluations, Qwen3.5-397B-A17B delivers aggressive efficiency throughout reasoning, agentic workflows, coding, and multimodal understanding. Here’s a take a look at its benchmarks and what it means:
Instruction Following & Reasoning
- IFBench (Instruction Following): 76.5 — among the many high scores in its class
- GPQA Diamond (Graduate-level reasoning): 88.4 — aggressive with frontier reasoning fashions
These outcomes recommend sturdy comprehension and structured reasoning that are essential for real-world workflows.
Agentic & Device Use Capabilities
- BFCL v4 (Agentic instrument use): 72.9
- BrowseComp (Agentic search): 78.6
- Terminal-Bench 2 (Agentic terminal coding): 52.5
Qwen 3.5 performs particularly effectively in agent-driven duties, reinforcing its positioning for workflow automation and power orchestration.
Coding & Developer Workflows
This locations it solidly within the vary of fashions able to dealing with actual coding and debugging workflows.
Multilingual Data
The rating aligns with its expanded language protection and improved information retrieval.
Multimodal & Visible Reasoning
- MMMU-Professional (Visible reasoning): 79.0
- OmniDocBench v1.5 (Doc understanding): 90.8
- Video-MME (Video reasoning): 87.5
- VITA-Bench (agentic multimodal interplay): 49.7
These numbers spotlight one in every of Qwen 3.5’s greatest strengths: multimodal comprehension throughout paperwork, visuals, and video.
Embodied & Spatial Reasoning
This displays bettering capabilities in real-world and embodied reasoning eventualities.
What These Benchmarks Actually Imply
As an alternative of dominating a single class, Qwen 3.5 reveals balanced energy throughout reasoning, agentic execution, coding, and multimodal understanding. That steadiness issues as a result of fashionable AI workloads aren’t single-task issues. They contain instruments, paperwork, photos, code, and multi-step workflows, and Qwen 3.5 seems to be constructed for precisely that actuality.
Palms-on With Qwen 3.5
We performed a few exams on each Qwen3.5 397B-A17B and the Qwen3.5-Plus. Listed here are the exams and the outcomes.
Process 1 – Coding with Qwen3.5-Plus
Immediate:
You might be an knowledgeable frontend developer and UI/UX designer.
Construct a contemporary, responsive promotional web site (single-page touchdown website) for the next occasion. The positioning needs to be visually premium, conversion-focused, and optimized for registrations.
Occasion Particulars:
Title: iqigai AI Fellowship Problem 2026
Tagline: India’s Largest AI and Knowledge Tech Hunt
Offered by: Fractal
Accomplice: Analytics Vidhya
Registration Hyperlink:
https://analyticsvidhya.com/datahack/contest/iqigai-genai-fellowship-challenge/?utm_source=social&utm_medium=X&utm_campaign=submitContent material to Embrace:
– Headline: India’s Largest AI and Knowledge Tech Hunt is now reside!
– Description:
The iqigai AI Fellowship Problem 2026 is greater than a hackathon — it’s a career-defining platform the place individuals compete, get nationally ranked, and achieve visibility amongst high employers.
– Dates: twentieth January – eighth March 2026
– Whole Prize Pool: ₹20 Lacs
– High Prizes:
Winner – ₹5 Lakhs
1st Runner-up – ₹3 Lakhs
2nd Runner-up – ₹2 LakhsWeb site Necessities:
1. Use HTML, CSS, and JavaScript (or React if most well-liked).
2. Totally responsive (desktop + cellular).
3. Fashionable gradient/AI-tech themed styling.
4. Easy scrolling navigation.
5. Clear CTA buttons linking to registration web page.
6. Sections:
– Hero part (massive headline + CTA)
– Concerning the Problem
– Key Highlights / Why Take part
– Prize Part (playing cards or visible badges)
– Timeline / Dates
– Name-to-Motion Banner
– FooterDesign Tips:
– Darkish tech gradient background
– Refined animations / hover results
– Clear typography
– Playing cards with shadows and rounded corners
– Non-compulsory icons or illustrations
– Preserve skilled occasion branding toneOutput Necessities:
– Present full runnable code
– Set up clearly into recordsdata
– Remark necessary elements
– Do NOT embody placeholder lorem ipsum
– Guarantee production-ready constructionGenerate the total web site code now.
Output:
Process 2 – Textual content-to-image with Qwen3.5-Plus
Immediate:
Create a cinematic anime-style transformation scene that includes Vegeta from Dragon Ball Tremendous unlocking Extremely Ego — depict a darkish cosmic battlefield as his physique radiates harmful god-like ki, muscle tissues 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 atmosphere; seize the essence of a God-of-Destruction mindset the place energy grows by way of battle depth and harm, emphasizing savage satisfaction, chaotic power waves, shattered terrain, and dramatic lighting — ultra-detailed, excessive distinction, dynamic digital camera angles, movement blur, and explosive anime shading, conveying overwhelming harmful dominance and unstoppable escalation.
Output:
Process 3 – Picture-to-video with Qwen3.5-Plus
Merely click on the Create Video choice on the Picture
Output:
Process 4 – Textual content-to-image with Qwen3.5 Open Weight
Immediate:
“Slash and Burn” may very well be a spirit or power of nature, embodying the cycle of destruction and renewal. It’d seem as a fiery, elemental being that consumes all the pieces in its path, just for new life to emerge from the ashes. This entity may very well be worshipped or feared as a deity of transformation and rebirth. backside left signature “sapope”
Output:
Process 5 – Picture-to-video with Qwen3.5 Open Weight
Merely click on the Create Video choice on the Picture
Output:
Ultimate Video:
Conclusion
The Qwen 3.5 household, with Qwen3.5 Open, is a step towards a extra succesful, unified AI system. With its hybrid MoE structure, native multimodal design, expanded language protection, and robust efficiency throughout reasoning, coding, and doc understanding benchmarks, Alibaba is clearly optimizing for real-world workloads.
What stands out most is the steadiness. As an alternative of excelling in a single slim activity, Qwen 3.5 reveals constant energy throughout agentic workflows, multimodal reasoning, and effectivity at scale. As AI strikes from chat interfaces to execution-driven techniques, fashions constructed for versatility and throughput will matter extra. With the benchmark performances and the outcomes we see in our hands-on exams, Qwen 3.5 positions itself firmly in that future.
Login to proceed studying and revel in expert-curated content material.
