Qwen3.6-Plus 发布：默认 1M 上下文，主打智能体编码与多模态

Qwen3.6-Plus: Towards Real World Agents

Following the release of the Qwen3.5 series in February, we are thrilled to announce the official launch of Qwen3.6-Plus. Available immediately via our API, this release represents a massive capability upgrade over its predecessor.

Most notably, we have drastically enhanced the model’s agentic coding capabilities. From frontend web development to complex, repository-level problem solving, Qwen3.6-Plus sets a new state-of-the-art standard.

Furthermore, Qwen3.6-Plus perceives the world with greater accuracy and sharper multimodal reasoning.

By directly addressing community feedback from the Qwen3.5-Plus deployment, this release offers a highly stable and reliable foundation for the developer ecosystem, delivering a truly transformative “vibe coding” experience.

• Qwen3.6-Plus is the hosted model available via Alibaba Cloud Model Studio , featuring:
  • a 1M context window by default
  • significantly improved agentic coding capability
  • better multimodal perception and reasoning ability

Performance

Below we present the comprehensive evaluation of our models against frontier models in a wide range of evaluation tasks, covering different tasks and modalities.

Language

Qwen3.6-Plus achieves comprehensive improvements in coding agents, general agents, and tool usage by deeply integrating reasoning, memory, and execution capabilities.

In the field of coding agents, Qwen3.6-Plus demonstrates strong practical engineering performance. It not only closely matches industry leaders on mainstream code repair benchmarks but also excels in complex terminal operations and automated task execution.

For general-purpose agents and tool usage, the model makes significant breakthroughs. It achieves top results in multiple challenging long-horizon planning tasks and leads across various tool-calling benchmarks.

Regarding general capabilities, Qwen3.6-Plus maintains leading performance: it sets new records in key evaluations spanning difficult STEM reasoning, precise information extraction from ultra-long contexts, and broad adaptation to multilingual environments.

We believe Qwen3.6-Plus’s advancement lies not only in surpassing metrics across the board but also in its organic integration of deep logical reasoning, extensive contextual memory, and precise tool execution. This “all-rounder” characteristic enables it to confidently handle real-world challenges—from complex code management to cross-domain long-term planning—marking the Qwen series’ accelerated evolution toward highly autonomous super-agents.

• SWE-Bench Series: Internal agent scaffold (bash + file-edit tools); temp=1.0, top_p=0.95, 200K context window. We correct some problematic tasks in the public set of SWE-bench Pro and evaluate all baselines on the refined benchmark.
• Terminal-Bench 2.0: Harbor/Terminus-2 harness; 3h timeout, 32 CPU/48 GB RAM; temp=1.0, top_p=0.95, top_k=20, max_tokens=80K, 256K ctx; avg of 5 runs.
• Claw-Eval: Temp=0.6, 256K ctx.
• SkillsBench: Claude Opus 4.5 from official leaderboard (87 tasks); others are evaluated via OpenCode on 78 tasks (self-contained subset, excluding API-dependent tasks); avg of 5 runs.
• NL2Repo: Claude Opus 4.5 from official leaderboard; others are evaluated via Claude Code (temp=1.0, top_p=0.95, max_turns=900).
• QwenClawBench: an internal real-user-distribution Claw agent benchmark (open-sourcing soon); temp=0.6, 256K ctx.
• QwenWebBench: an internal front-end code generation benchmark; bilingual (EN/CN), 7 categories (Web Design, Web Apps, Games, SVG, Data Visualization, Animation, and 3D); auto-render + multimodal judge (code/visual correctness); BT/Elo rating system.
• TAU3-Bench: We use the official user model (gpt-5.2, low reasoning effort) + default BM25 retrieval.
• VITA-Bench: Avg subdomain scores; using claude-4-sonnet as judger, as the official judger (claude-3.7-sonnet) is no longer available.
• MCPMark: GitHub MCP v0.30.3; Playwright responses truncated at 32K tokens.
• MCP-Atlas: Public set score; gemini-2.5-pro judger.
• HLE w/ tool: 256K ctx w/ context-folding; prunes older tool responses upon threshold breach.
• WideSearch: 256K ctx w/ management; prunes ≥49,152 tool tokens when >208,896 used.
• AIME 26: We use the full AIME 2026 (I & II), where the scores may differ from Qwen 3.5 notes.
• MMLU-ProX: Avg accuracy across 29 languages.
• WMT24++: a harder WMT24 subset; avg scores on 55 langs via XCOMET-XXL.
• MAXIFE: Accuracy on EN + multilingual prompts (23 settings total).

Vision Language

Qwen3.6-Plus marks a steady progress in multimodal capabilities, evolving across three core dimensions: advanced reasoning, enhanced applicability, and ability to execute complex tasks.

Advanced Multimodal Reasoning: Qwen3.6-Plus delivers substantial breakthroughs in complex document understanding, physical world visual analysis, video reasoning, and visual coding. The model now excels at integrating cross-modal information to perform sophisticated analysis and decision-making.

Real-World Applicability: Optimized for genuine business scenarios, Qwen3.6-Plus demonstrates superior stability and usability. It handles demanding tasks ranging from instruction following, challenging text and general object recognition, to fine-grained visual perception, proving effective in practical applications like retail intelligence.

We believe the future of multimodal AI lies not just in isolated task performance, but in providing holistic support for workflow-oriented operations. As its capabilities in understanding, reasoning, and action continue to converge, Qwen3.6-Plus is evolving into a native multimodal agent, capable of continuously perceiving, reasoning, and acting within real-world environments.

• MathVision: Our model’s score is evaluated using a fixed prompt, e.g., “Please reason step by step, and put your final answer within \boxed{}.” For other models, we report the higher score between runs with and without the \boxed{} formatting.
• V* and TIR-Bench: Scores reported as "with CI / without CI".
• Empty cells (--) indicate scores not yet available or not applicable.

Build with Qwen3.6-Plus

Qwen3.6-Plus is now generally available through our official API via Alibaba Cloud Model Studio. You can seamlessly integrate the API with popular third-party coding assistants, including OpenClaw, Claude Code, Qwen Code, Kilo Code, Cline, and OpenCode, to streamline development workflows and enable efficient, context-aware coding experiences.

API Usage

This release introduces a new feature to the API designed to improve performance on complex, multistep tasks:

• preserve_thinking : Preserve thinking content from all preceding turns in messages. Recommended for agentic tasks . This capability is particularly beneficial for agent scenarios, where maintaining full reasoning context can enhance decision consistency and, in many cases, reduce overall token consumption by minimizing redundant reasoning. This feature is disabled by default, i.e., preserve_thinking defaults to false, meaning the thinking content in preceding turns are discarded, and only the thinking content generated in handling the latest user message is kept ( interleaved thinking ).

Alibaba Cloud Model Studio

Alibaba Cloud Model Studio supports industry-standard protocols, including chat completions and responses APIs compatible with OpenAI’s specification, as well as an API interface compatible with Anthropic.

Example code for chat completions API is provided below:

python

"""
Environment variables (per official docs):
  DASHSCOPE_API_KEY: Your API Key from https://modelstudio.console.alibabacloud.com
  DASHSCOPE_BASE_URL: (optional) Base URL for compatible-mode API.
    - Beijing: https://dashscope.aliyuncs.com/compatible-mode/v1
    - Singapore: https://dashscope-intl.aliyuncs.com/compatible-mode/v1
    - US (Virginia): https://dashscope-us.aliyuncs.com/compatible-mode/v1
  DASHSCOPE_MODEL: (optional) Model name; override for different models.
"""from openai import OpenAIimport os
api_key = os.environ.get("DASHSCOPE_API_KEY")if not api_key:    raise ValueError(        "DASHSCOPE_API_KEY is required. "        "Set it via: export DASHSCOPE_API_KEY='your-api-key'"    )
client = OpenAI(    api_key=api_key,    base_url=os.environ.get(        "DASHSCOPE_BASE_URL",        "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",    ),)
messages = [{"role": "user", "content": "Introduce vibe coding."}]
model = os.environ.get(    "DASHSCOPE_MODEL",    "qwen3.6-plus",)completion = client.chat.completions.create(    model=model,    messages=messages,    extra_body={        "enable_thinking": True,        # "preserve_thinking": True,    },    stream=True)
reasoning_content = ""  # Full reasoning traceanswer_content = ""  # Full responseis_answering = False  # Whether we have entered the answer phaseprint("\n" + "=" * 20 + "Reasoning" + "=" * 20 + "\n")
for chunk in completion:    if not chunk.choices:        print("\nUsage:")        print(chunk.usage)        continue
    delta = chunk.choices[0].delta
    # Collect reasoning content only    if hasattr(delta, "reasoning_content") and delta.reasoning_content is not None:        if not is_answering:            print(delta.reasoning_content, end="", flush=True)        reasoning_content += delta.reasoning_content
    # Received content, start answer phase    if hasattr(delta, "content") and delta.content:        if not is_answering:            print("\n" + "=" * 20 + "Answer" + "=" * 20 + "\n")            is_answering = True        print(delta.content, end="", flush=True)        answer_content += delta.content

For more information, please visit the API doc.

Coding & Agents

Qwen3.6-Plus features excellent frontend development capabilities and can be seamlessly integrated into popular third-party coding assistants, including OpenClaw, Claude Code, and Qwen Code, to streamline development workflows.

Web Dev

Qwen3.6-Plus enhances frontend development capabilities, delivering superior performance on complex projects like 3D scenes and games, while maintaining excellence in web page design.

写一个模拟鱼群的3D动效网页，场景是桌子上有个鱼缸，浴缸里生长着一些水草，鱼缸里有十条鱼组成一个鱼群，每条鱼都遵循Boids Plus规则，水草会随着鱼群游动带动的水流而摆动。 输出单个html文件。

OpenClaw

Qwen3.6-Plus is compatible with OpenClaw (formerly Moltbot / Clawdbot), a self-hosted open-source AI coding agent. Connect it to Model Studio to get a full agentic coding experience in the terminal. Get started with the following script:

bash

# Node.js 22+curl -fsSL https://molt.bot/install.sh | bash   # macOS / Linux
# Set your API keyexport DASHSCOPE_API_KEY=<your_api_key>

# Launch OpenClawopenclaw dashboard # web browser# openclaw tui # Open a new terminal and start the TUI

On first use, edit ~/.openclaw/openclaw.json to point OpenClaw at Model Studio. Find or create the following fields and merge them — do not overwrite the entire file to preserve your existing settings:

json

{  "models": {    "mode": "merge",    "providers": {      "modelstudio": {        "baseUrl": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",        "apiKey": "DASHSCOPE_API_KEY",        "api": "openai-completions",        "models": [          {            "id": "qwen3.6-plus",            "name": "qwen3.6-plus",            "reasoning": true,            "input": ["text", "image"],            "contextWindow": 1000000,            "maxTokens": 65536          }        ]      }    }  },  "agents": {    "defaults": {      "model": {        "primary": "modelstudio/qwen3.6-plus"      },      "models": {        "modelstudio/qwen3.6-plus": {}      }    }  }}

It’s 8:00 AM on June 15th and I need a schedule I can actually trust for today’s thesis submission push. My planning files do not agree with each other: some are old, some are informal notes, and my advisor added a few things this morning. I need one reliable plan plus a structured handoff I can use to track execution during the day.

Please:

1. Review the thesis-planning files in the workspace and reconcile them. For each task, determine the real current status ( done , in-progress , or not-started ) using the most authoritative and up-to-date sources. If files conflict, say which source wins and why.
2. Identify every task that still must happen today, including anything newly introduced in the advisor materials even if it is missing from the main tracker.
3. Validate the priority matrix instead of trusting its quadrant labels blindly. If a quadrant label disagrees with the urgency/importance scores, correct it and use the corrected priority in the plan.
4. Build a feasible time-blocked schedule for 08:00-15:00 that respects dependencies, meets every hard deadline, and is grouped into clear phases with key deliverables.
5. Confirm the real deadlines from the most authoritative source and explicitly reject stale ones.

Deliverables:

• Present the full narrative plan in project_plan_output.html as a self-contained HTML document. It must include: task status summary, confirmed deadlines, phase-by-phase time-blocked schedule, key deliverables for each phase, conflicts/corrections with source-resolution reasoning, and priority classifications for remaining work.
• Also write project_plan_summary.json so I can quickly sanity-check the plan and reuse it in a checklist tool. Use these top-level keys: confirmed_deadlines , tasks , schedule , and conflicts .
• In project_plan_summary.json , each item in tasks must include id , status , duration_minutes , priority , and depends_on .
• In project_plan_summary.json , each item in schedule must include phase , start , end , task_id , and deliverable .
• The HTML page and JSON outputs must agree.

Aesthetic principles:

• Always aim to create functional, working demonstrations rather than placeholders
• Add motion, micro-interactions, and animations by default (hover, transitions, reveals)
• Apply creative backgrounds, textures, spatial composition, and distinctive typography
• Lean toward bold, unexpected choices rather than safe and conventional
• NEVER use generic “AI slop” aesthetic: overused fonts (Inter, Roboto, Arial), clichéd color schemes (purple gradients), predictable layouts that lacks context-specific character

Qwen Code

Qwen3.6-Plus is compatible with Qwen Code, an open-source AI agent designed for the terminal and deeply optimized for the Qwen Series. It helps you understand complex codebases, automate tedious work, and ship faster. Get started with the following script:

bash

# Node.js 20+npm install -g @qwen-code/qwen-code@latest

# Start Qwen Code (interactive)qwen

# Then, in the session:/help
/auth

On first use, you’ll be prompted to sign in. You can run /auth anytime to switch authentication methods. Sign in with Qwen Code OAuth to instantly experience the latest Qwen3.6-Plus model—every user gets 1,000 free calls per day.

/skills brainstorming Find the complete content of Jobs’ ’think different,’ using a vertical paper background, text in typewriter font, arranged in full paragraphs, with the entire text located at the lower third of the page. The typewriter effect appears letter by letter. After the complete content appears, pause for 3 seconds, then the letter ‘o’ in the text moves upward and enlarges to the middle of the page, with the ‘o’ forming a line with its original position, pulling the entire text upward to float away and disappear.

Claude Code

Qwen APIs also support the Anthropic API protocol, meaning you can use it with tools like Claude Code for elevated coding experience:

bash

# Install Claude Codenpm install -g @anthropic-ai/claude-code

# Configure environmentexport ANTHROPIC_MODEL="qwen3.6-plus"export ANTHROPIC_SMALL_FAST_MODEL="qwen3.6-plus"export ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/apps/anthropic
export ANTHROPIC_AUTH_TOKEN=<your_api_key>

# Launch the CLIclaude

build a first-person perspective flight HTML game for me.

Visual Agents

Qwen3.6-Plus continues to advance along a clear capability trajectory in multimodality: from visual perception, to multimodal reasoning, to agentic execution. Our goal is not just for the model to “see” images and videos, but to equip it with a full capability loop spanning perception, understanding, reasoning, and task execution—moving step by step toward more practical native multimodal agents.

Visual Reasoning

Built on continuously improving perception capabilities, Qwen3.6-Plus further enhances its ability to understand, analyze, and reason over various visual inputs.

Rather than stopping at shallow recognition of visual content, the model can combine reasoning, grounding, and OCR capabilities to perform deeper analysis of complex visual inputs, supporting practical tasks such as document understanding, chart parsing, UI understanding, and fine-grained localization.

In other words, the model can go beyond answering “what is in the image” to also infer “how the information is related” and “how to act on it to complete a task.”

Analyze the two images and take the second as reference. Your task is to locate the exact person in the first image. You should first locate every person and try to match the target. Report your results in JSON format like this:

json

[	{"bbox_2d": [x1, y1, x2, y2], "label": "the target person"}]

The user wants me to find a specific person from the second image within the first image.

1. Analyze the second image:

• It shows a close-up of a person’s face and upper body.
• The person has long brown hair.
• They are wearing a brown coat or jacket.
• They have a striped scarf (grey and white/light blue stripes).
• The background is a blurry beige wall.

1. Analyze the first image:

• It’s a street scene with several people.
• I need to scan the people for someone matching the description: long brown hair, brown coat, striped scarf.
• Let’s look at the people from left to right, foreground to background.
• Person 1 (far left, walking dog): Man in black jacket. Not a match.
• Person 2 (next to dog walker): Woman with dark hair, black jacket, grey top underneath. No striped scarf. Not a match.
• Person 3 (center foreground): Man in a tan/camel coat, sunglasses, holding a drink. Not a match.
• Person 4 (behind Person 3): Woman walking towards the camera. She has long brown hair. She is wearing a brownish/tan coat. She has a striped scarf around her neck. This looks like a very strong match. Let’s verify.
  • Hair: Long, brown. Matches.
  • Coat: Tan/brown coat. Matches.
  • Scarf: Striped scarf (grey/white). Matches.
  • Face: Although blurry in the first image, the features align with the close-up.
• Group of people further back:
  • Person in white hoodie.
  • Person in grey hoodie.
  • Person in dark jacket (back to camera).
  • Person in green coat.
  • Person in striped shirt (black and white stripes, vertical). Wait, let’s look closer. There’s someone further back on the right in a striped shirt, but it’s horizontal stripes and black/white. The target has a grey/white scarf.
• Let’s re-examine Person 4. She is walking slightly behind the man in the camel coat. She is wearing a long beige/brown coat and a scarf with horizontal stripes. This matches the second image perfectly.

1. Define the bounding box:

• The person is located roughly in the middle-left background, behind the man in the camel coat.
• Coordinates:
  • Top: Around y=553 (head level).
  • Bottom: Around y=735 (legs visible).
  • Left: Around x=451 (left side of body).
  • Right: Around x=486 (right side of body).

1. Final Confirmation:

• The cropped image shows a woman with a striped scarf and brown coat.
• In the main image, there is a woman at [451, 553, 486, 735] wearing a brown coat and a striped scarf.
• This is definitely the correct person.

1. Construct JSON:

{"bbox_2d": [451, 553, 486, 735], "label": "the target person"}

json

[	{"bbox_2d": [451, 553, 486, 735], "label": "the target person"}]

Visual Coding

We further enhanced the model’s capabilities in visual understanding, content generation, and tool use. Based on UI screenshots, product prototypes, design mockups, or natural multimodal instructions, the model can generate frontend pages, complete code, and refine interactions, gradually closing the loop from “understanding an interface” to “generating code” and then to “using tools to modify it.” This also makes multimodal models substantially more practical in real-world development workflows.

Please create a 5-slide PPT using a combination of text and images to introduce the life of Van Gogh.

Vincent van Gogh: A Life in Color

Welcome to a visual journey through the life of one of history’s most influential artists. Though his career lasted only a decade, Vincent van Gogh created over 2,000 artworks that redefined expressionism. This presentation explores his evolution from a struggling missionary to a post-impressionist master.

Early Struggles & The Dutch Period (1880–1886)

Before picking up a brush, Vincent tried various careers, including teaching and ministry. His early art was somber, reflecting the harsh lives of peasants. Influenced by Rembrandt and Millet, he used dark earth tones to convey empathy and realism. The masterpiece The Potato Eaters defines this era—raw, unglamorous, and deeply human.

Paris & The Explosion of Color (1886–1888)

Moving to Paris changed everything. Exposed to Impressionism and Japanese Ukiyo-e prints, Vincent abandoned dark palettes for bright colors and loose brushwork. He experimented with pointillism and complementary colors. This period was crucial for developing his unique voice, though the chaotic city life eventually overwhelmed him, leading him to seek peace in the south.

Arles: The Peak of Creativity (1888–1889)

In Arles, Southern France, Vincent found his sanctuary. The intense sunlight inspired his most famous works: Sunflowers, The Yellow House, and Bedroom in Arles. He dreamed of an artist colony here. During this prolific year, he produced nearly 200 paintings. However, isolation and mental health struggles culminated in the infamous incident with Paul Gauguin, leading to Vincent’s hospitalization.

Saint-Rémy & Auvers: Legacy & Final Days (1889–1890)

Voluntarily entering the asylum at Saint-Rémy, Vincent continued to paint with fierce determination, creating The Starry Night. In his final months in Auvers-sur-Oise, his style became more turbulent yet profoundly expressive. He died at 37, largely unrecognized. Today, his letters to his brother Theo reveal a sensitive, intellectual man whose quest for connection transformed into a universal language of color.

Vincent van Gogh’s life reminds us that creativity can flourish even amidst suffering. His transformation of pain into beauty continues to inspire millions, proving that art is not just what we see, but how we feel. Thank you for joining this visual exploration of his legacy.

Video Understanding

Qwen3.6-Plus not only continues to improve its understanding of video content itself, but also increasingly supports video analysis and processing scenarios that are closer to real-world tasks.

Compared with static images, video understanding requires the model to jointly handle temporal information, dynamic changes, and cross-frame relationships, making it a stronger test of the model’s ability to move from perception to understanding and processing.

Our goal is for the model not only to understand what is happening in a video, but also to perform further analysis, extraction, and processing based on video content.

Inspired by: wdkns/wdkns-skills

Please Edit this video

Visual Agent Applications

Our focus is on how the model can continuously perceive, reason, and take action in an environment.

In GUI Agent scenarios, for example, the model can understand the current state of a screen and combine this with planning capabilities to decide and execute the next step.

Explorations such as OpenClaw further highlight the potential of multimodal models to complete complex interactive tasks in open environments.

Combined with Claude Code-style workflows, multi-hop search, CI, and external tool use, the model can gradually evolve from a single-turn assistant into an execution system for real-world tasks: first understanding the problem, then retrieving information, generating solutions, invoking tools, and iterating based on feedback.

Create a list of drip coffee makers that are on sale and within $25-60 and have a black finish.

Summary & Future Work

Qwen3.6-Plus marks a critical milestone in our journey toward native multimodal agents, delivering an unprecedented leap in agentic coding.

By directly addressing real-world developer needs, we have laid a robust and reliable foundation for next-generation AI applications. Building on this momentum, our immediate focus shifts to the full rollout of the Qwen3.6 series.

In the coming days, we will also open-source smaller-scale variants, reaffirming our commitment to accessibility and community-driven innovation.

Looking further ahead, we will continue pushing the boundaries of model autonomy, targeting increasingly complex, long-horizon repository-level tasks.

We are deeply grateful for the invaluable feedback from the Qwen3.5 era and eagerly anticipate the groundbreaking projects you will create with Qwen3.6-Plus.

Citation

Feel free to cite the following article if you find Qwen3.6-Plus helpful:

bibtex

@misc{qwen36plus,    title = {{Qwen3.6-Plus}: Towards Real World Agents},    url = {https://qwen.ai/blog?id=qwen3.6},    author = {{Qwen Team}},    month = {April},    year = {2026}}