On April 2, 2026, while China’s tech community was still abuzz with rumors of DeepSeek V4, Alibaba quietly launched a blockbuster move. With no grand launch event or overwhelming pre-heating, Qwen 3.6-Plus made its low-key debut on Alibaba Cloud’s Bailian Platform, yet sent shockwaves across the entire AI industry.

Within just 24 hours of launch, it surged to the top of the daily rankings on OpenRouter, a world-renowned large model API call platform, with single-day token calls exceeding 1.4 trillion, directly setting a new global record for single-model daily call volume on the platform. OpenRouter officials even gave it high praise, calling it the “strongest performance of any new model in history”.

Even more stunning is its breakthrough in programming capabilities. On the Code Arena rankings under LMArena, which focuses on AI programming, Qwen 3.6-Plus secured second place globally, outperforming international giants such as OpenAI, Google, and xAI to become the highest-ranked Chinese large model on the list. Notably, it scored 1,452 points in the React specialized ranking, trailing only Claude Opus 4.6 Thinking (1,540 points) by 88 points, surpassing OpenAI’s newly released GPT-5.0-High (1,448 points) by 4 points, and leading Google Gemini 3.1 Pro Preview (1,440 points) by 12 points.

This means that in the most challenging AI Coding and Agent tasks, Qwen 3.6-Plus boasts code generation and engineering capabilities comparable to, and even surpassing, the world’s top large models. With this impressive achievement, Alibaba jumped to fourth place in the global AI lab rankings, behind only Anthropic, OpenAI, and Google, rewriting its own AI development narrative overnight.

I. This Is No Ordinary Upgrade: Agent Capability Is the Real Game-Changer

Judging solely by parameter scale, Qwen 3.6-Plus may seem unremarkable—its parameters are less than half those of Kimi K2.5 or GLM-5. However, the truly disruptive upgrade lies in its core “Agent” capability, which completely breaks the limitations of traditional AI.

1.1 From “Writing Code” to “Getting Things Done”: AI Moves Beyond the “Intern” Role

Traditional AI was more like an inflexible “intern”: it would output code upon receiving instructions, but never checked if the code ran properly, let alone fix errors autonomously. Humans had to debug and provide feedback repeatedly throughout the process, resulting in low efficiency.

Qwen 3.6-Plus introduces the Agentic Coding paradigm, achieving a qualitative leap from “outputting code” to “completing tasks”, with five core capabilities:

• Autonomous planning (disassembling requirements and mapping execution paths after understanding needs)
• Tool invocation (automatically calling editors, terminals, shell commands, etc.)
• Execution verification (running code autonomously to validate results without human intervention)
• Automatic repair (identifying error logs, locating issues, and submitting fixes)
• Long-range task planning (breaking down complex requirements, executing step-by-step, and continuously verifying)

Real-world test cases have stunned the industry: when prompted to “build a responsive corporate official website”, the model automatically generated a complete HTML/CSS/JS project in just 8 minutes, complete with image placeholders, navigation bars, contact forms, and even auto-deployed preview links—all at a total cost of only 0.15 yuan.

In real-world agent benchmarks such as Claw-Eval and QwenClawBench, Qwen 3.6-Plus boosted its overall task completion rate by 10%–20%, with a success rate of over 70% for complex tasks. In contrast, traditional models typically hover around a 50% success rate, marking a stark gap.

1.2 ATH Architecture: Teaching AI to “Think While Acting” with a Built-In Self-Check Mechanism

Powered by the ATH (Agentic-Task-Hybrid) architecture, Qwen 3.6-Plus features a built-in “self-checking loop” at its core. Instead of delivering code immediately after generation, it first runs a trial execution within its own inference space. If errors occur, it autonomously analyzes causes, fixes the code, and only submits a ready-to-run final product to the user.

As one developer put it: “It no longer chases ‘optimal single output’; it is designed around ‘getting a job done fully’. It acts more like an experienced engineer than a mere code-writing tool.”

This shift in thinking elevates the model from a one-shot “input-output” mapping to a complete “understand-plan-execute-correct” workflow, often involving 3 to 8 or more decision steps, completely solving the traditional AI pain point of “writing but not fixing code”.

Additionally, Qwen 3.6-Plus is fully compatible with six mainstream Agent frameworks: OpenClaw, Qwen Code, Claude Code, KiloCode, Cline, and OpenCode. Developers can integrate the model into existing “Lobster” workflows to complete the entire planning-to-execution process in the terminal. Even more conveniently, Alibaba has opened access to Qwen 3.6-Plus via the Anthropic API protocol. With minimal configuration, developers can directly point existing Claude Code setups to the new model for a seamless migration experience.

II. 1M Context Window: The Superpower to “Swallow” Entire Code Repositories

A 1,000,000-token default context window—equivalent to roughly 750,000 Chinese characters—allows Qwen 3.6-Plus to ingest an entire codebase, hundreds of pages of legal documents, a full year of product requirement documents, or even the complete Three-Body trilogy in one go. Its maximum output length reaches 65,536 tokens, enough to generate full project architecture documents or large code modules in a single pass.

Traditional models suffer severe performance degradation in their attention mechanisms when processing ultra-large contexts. However, Qwen 3.6-Plus’s native long-context capability, paired with an efficient hybrid sparse Mixture of Experts (MoE) architecture (approximately 397B total parameters, with only ~17B activated during inference), makes ultra-large context processing practical, solving the longstanding pain point of traditional models “failing to retain details”.

This capability delivers exceptional value in real-world scenarios:

• Scenario 1: Code Review – Import an entire company’s code repository to analyze architectural decisions, locate cross-file bugs, and assess technical debt, saving massive hours of manual review.
• Scenario 2: Technical Document Q&A – Upload hundreds of pages of technical manuals; AI quickly locates specific concepts and answers questions without manual browsing.
• Scenario 3: Multi-Round Complex Dialogue – The 1M token window covers lengthy conversation histories, eliminating mid-dialogue “memory lapses” and resolving long-range communication pain points.

One reviewer commented: “You can feed in an entire project’s codebase, a dozen PRDs, or even hundreds of pages of UI design specs. It remembers every detail, no more ‘fixing one thing and breaking another’.”

III. Programming Dominance: Outperforming Rivals by Multiple Times with Half the Parameters

3.1 Benchmark Tests Match Top Models, Surpass in Some Metrics

Qwen 3.6-Plus delivered a convincing performance across multiple authoritative programming benchmarks, matching Claude Opus 4.5 and even outperforming it in some indicators:

表格

Source: Public benchmark tests and technical reports

The 61.6 score on Terminal-Bench 2.0 is particularly outstanding—Qwen 3.6-Plus is the only model to top this benchmark, which evaluates terminal operation and automated task execution capabilities, the core of Agentic Coding implementation.

In the Code Arena blind test rankings, its 1,452-point React specialized score ranked second globally, surpassing OpenAI GPT-5.0-High and Google Gemini 3.1 Pro Preview and demonstrating strong programming competitiveness. What’s more remarkable is that with less than half the parameters of Kimi K2.5 or GLM-5, it matches or outperforms them in performance, highlighting its efficiency advantage.

3.2 “Vibe Coding” Realized: Bring Ideas to Life Without Coding Knowledge

Alongside its leap in programming capabilities, the concept of “Vibe Coding” has finally moved from vision to reality.

You need no coding skills—just aesthetics and ideas. For example, feed a sketch to the AI and say: “I want a The Legend of Zelda vibe, clear skies, crisp snow-capped mountains, glistening falling snow, and WASD controls for camera rotation.” In under a minute, Qwen 3.6-Plus generates a browser-run 3D scene with auto-tuned snow gravity, wind drift, and smooth camera movement.

This is the core of Vibe Coding: not “implementing ideas with code”, but “driving code with intuition”. Unlike the repetitive “write code-error-feedback-revise” cycle of traditional AI, Qwen 3.6-Plus debugs, fixes, and runs code autonomously, delivering ready-to-use final products to users and drastically lowering the barrier to programming.

IV. Shift to Closed-Source: Is the Open-Source Era Truly Over?

4.1 Behind Closed-Source: A Pragmatic Commercial Choice

Unlike the previously open-source Qwen series, Qwen 3.6-Plus adopts a proprietary model strategy, with no weight downloads available and services provided only via API. The news sparked heated debate in the AI community: some developers complained that “Qwen is following OpenAI’s closed-source money-making model”, while others lamented “the end of the open-source era”.

However, Alibaba provided a clear explanation: smaller-parameter models of the Qwen 3.6 series will be open-sourced later, with only the Plus version remaining closed-source. This shift is rooted in highly pragmatic industry logic:

1. Sustainable commercialization: Training and inference costs for large models are extremely high. Closed-source APIs are currently the most sustainable monetization path and the foundation for continuous model iteration. An Alibaba Cloud spokesperson noted this aligns with industry trends— as cutting-edge models scale up, local hardware deployment becomes increasingly impractical, pushing enterprises to monetize traffic via official cloud platforms.
2. Guaranteed service stability: Closed-source allows Alibaba to optimize and monitor APIs end-to-end, providing stable service SLAs for enterprise users and avoiding compatibility and performance issues developers face with local hardware deployment.
3. Building a data closed loop: User usage data collected via APIs helps Alibaba continuously optimize the model, creating a positive flywheel: “more users → better model → more users” for a virtuous cycle.

Today, among China’s leading open-source large model enterprises, most have shifted to closed-source—with DeepSeek likely to remain open-source and Kimi’s stance unknown. Qwen’s closed-source transition is not an isolated case, but an inevitable choice as the AI industry moves from “technical exploration” to “commercial implementation”. The divergent paths of DeepSeek and Alibaba will now spark a new market rivalry.

4.2 Price War Ignited: Costs Just One-Ninth of Claude’s

Qwen 3.6-Plus’s API pricing is a knockout blow to closed-source competitors, drastically reducing costs for developers and enterprises. The specific pricing is as follows:

表格

Source: Official Alibaba Cloud pricing

For comparison, Claude API’s output price is as high as $15 per million tokens (≈¥109), meaning Qwen 3.6-Plus’s output cost is just one-ninth of Claude’s, with an outstanding cost-performance ratio.

The 70-million-token free quota for new users is particularly generous: a single developer can complete roughly 400 8-minute official website generation tasks or thousands of code generation and debugging sessions entirely for free, greatly lowering the trial barrier.

4.3 Three Launches in 72 Hours: Alibaba AI’s “Muscle Flexing”

The release of Qwen 3.6-Plus was the finale of three blockbuster models launched by Alibaba in 72 hours, signaling accelerated AI strategic deployment:

• March 30: Qwen 3.5-Omni, a multimodal interactive model, outperforming Gemini-3.1 Pro in multiple metrics across 215 audio-visual understanding tasks.
• April 1: Wan2.7-Image, a text-to-image model, ranking first in China in human preference blind tests for Chinese text-to-image generation.
• April 2: Qwen 3.6-Plus, a flagship programming and Agent model, topping OpenRouter’s daily rankings and securing second place globally on Code Arena.

This intensive launch came just two weeks after Alibaba established the ATH (Alibaba Token Hub) Business Group. The group integrates the computing power layer (Alibaba Cloud), model layer (Tongyi), and application layer (Wukong, Qwen App, etc.) to build a complete closed loop of “create tokens – transport tokens – consume tokens”. The three model releases all align with a single strategic focus: Model as an Agent—transforming AI from a tool into a “digital employee” capable of autonomously completing tasks.

V. Starlink Engine: An Efficient API Aggregation Solution, Even Better with 4SAPI

While unpacking Qwen 3.6-Plus’s powerful capabilities, I’d also like to recommend an indispensable tool in my daily work: the Starlink Engine. It perfectly solves the hassle of multi-model API calls and saves me substantial API subscription costs.

If you want to call all mainstream models with a single API Key, or need a secure, stable, direct domestic API solution, the Starlink Engine is well worth trying. Pairing it with 4SAPI (4SAPI.COM) elevates the experience even further: as an enterprise-grade unified access platform for large model APIs, 4SAPI is compatible with the OpenAI interface protocol, supporting zero-cost adaptation to the Starlink Engine and over 650 mainstream models including Qwen 3.6-Plus, Claude, and GPT. Model switching takes just one line of code without tedious adaptation, further cutting development costs and maximizing the Starlink Engine’s advantages.

VI. Qwen 3.6-Max Ready to Launch: An All-Round Flagship Beyond the Plus Version

Notably, Qwen 3.6-Plus is only the first model in the Qwen 3.6 series. The highly anticipated flagship model Qwen 3.6-Max has been officially announced for an upcoming release.

This means Qwen’s complete large model product matrix is taking shape at an accelerated pace:

• Plus Version: A production-grade API for developers and enterprises, focusing on core programming and Agent capabilities with exceptional cost-performance.
• Max Version (Coming Soon): A more powerful all-round flagship, expected to achieve further breakthroughs in multimodality and general reasoning, competing with the world’s top all-round models.
• Lightweight Open-Source Version (Later Release): Small-parameter open-source models, upholding Alibaba’s commitment to the developer community and suitable for local deployment.

Meanwhile, Qwen 3.6-Plus has been fully integrated into Alibaba’s internal productivity system—including the AI-native enterprise platform Wukong, the Qwen App, and the programming tool Qoder. This means Qwen large models are no longer isolated chat tools, but core infrastructure on the supply side of Alibaba’s entire AI business.

VII. When AI Evolves from “Toy” to “Colleague”: Ordinary People Reap Tech Dividends

The launch of Qwen 3.6-Plus sends a clear signal: the AI industry’s competitive logic is shifting from “competing on parameters and scores” to “competing on practicality and cost”.

Over the past two years, the industry fell into the trap of a “parameter race”, with players vying for larger model parameters and higher benchmark scores. Qwen 3.6-Plus, however, chose a different path: using fewer parameters, lower costs, and stronger Agent capabilities to make AI truly “work for real-world tasks”, turning it from a “luxury” into a “daily necessity”.

A product manager with no coding skills can build a small game with 3D scenes in a day via Vibe Coding. A startup team can set up a fully automated pipeline from requirement analysis to code delivery using Qwen 3.6-Plus and Agent frameworks, without hiring a large engineering team. A senior developer can access near-Claude-level programming capabilities for just tens of yuan in API costs monthly, no longer locked into high-priced APIs by closed-source giants.

Of course, Qwen 3.6-Plus is not perfect. Independent reviewers have found that while the model runs error-free, it has limitations in handling complex business logic, and its group chat role function fails to fully meet design goals in specific scenarios. This shows Chinese models still have room to improve on their journey to becoming “the world’s best”.

Yet this is a positive development: fiercer industry competition drives faster technological iteration, and the ultimate beneficiaries will always be developers and ordinary people. When AI truly evolves from a “toy” to a “colleague”, we can shed tedious repetitive work and focus on more creative endeavors—this is perhaps the core value of AI technology.