[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-qwen38-max-agentic-work-real-frontier-en":3,"article-related-qwen38-max-agentic-work-real-frontier-en":30,"series-model-release-ff8312eb-e6ac-4bbc-85d5-38844a1c1964":75},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":23,"views":27,"created_at":28,"published_at":29,"topic_cluster_id":11},"ff8312eb-e6ac-4bbc-85d5-38844a1c1964","qwen38-max-agentic-work-real-frontier-en","Qwen3.8-Max proves that agentic work is the real frontier","\u003Cp data-speakable=\"summary\">2.4 trillion parameters and 1 million tokens make Qwen3.8-Max a serious agentic work model.\u003C\u002Fp>\u003Cp>Qwen3.8-Max is the kind of release that changes how teams should think about frontier models: not as clever chat systems, but as systems that can actually finish work. Alibaba says the model scales to 2.4 trillion parameters, activates 95B parameters, and handles 1 million-\u003Ca href=\"\u002Ftag\u002Ftoken\">token\u003C\u002Fa> contexts, while also improving across coding, office work, research, and long-horizon tasks. That is not a cosmetic upgrade. It is a statement that the next phase of model competition is no longer about isolated prompts or single-turn \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> glory, but about whether a model can stay useful across days of work, changing tools, and messy real-world feedback.\u003C\u002Fp>\u003Ch2>First argument: breadth matters more than one-off benchmark wins\u003C\u002Fh2>\u003Cp>The strongest evidence is the coding and research traces. In one case, Qwen3.8-Max was tasked with building an auto-evolving harness project from an empty folder and worked for roughly 16 days with no human help, producing 265 commits, 127 pull requests, and 151 issues. In another, it spent about five days reproducing and then improving a recent reasoning paper, writing around 7,600 lines of code, executing more than 1,100 steps, and running 33 GPU training rounds. Those are not toy demos. They show a model that can sustain intent, repair its own mistakes, and keep moving toward a deliverable when the task is no longer a single answer but a living codebase or research loop.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785844979892-cdc1.png\" alt=\"Qwen3.8-Max proves that agentic work is the real frontier\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>That matters because the market has already moved past “can it write a function?” The real question is whether the model can own a project slice from start to finish. A model that can scaffold an experiment, debug the pipeline, rerun the training, and extract a better result is more valuable than one that merely scores well on a static benchmark. Qwen3.8-Max’s reported gains on Text Arena, CodeArena, and VisionArena fit that pattern: the release is trying to prove competence in environments where the work itself is dynamic, not frozen into a test set.\u003C\u002Fp>\u003Ch2>Second argument: the training recipe is the product\u003C\u002Fh2>\u003Cp>Qwen’s office-work story is especially important because it explains why the model generalizes. The team says it expanded real-world RL by separating environment complexity into three axes: task length, workspace complexity, and toolchain variety. It also unified reward across execution checks, text evaluation, visual inspection, and behavior auditing, then balanced batches so training signal stayed stable across task types. That is a serious design choice. It means the model is not being optimized for one narrow tool or one narrow benchmark style, but for the harder problem of remaining reliable as the workspace changes under it.\u003C\u002Fp>\u003Cp>The practical payoff is visible in the harness results. Qwen3.8-Max is reported to perform strongly across QwenWork, \u003Ca href=\"\u002Fnews\u002Fprompt-engineering-overrated-claude-code-en\">Claude Code\u003C\u002Fa>, \u003Ca href=\"\u002Ftag\u002Fcodex\">Codex\u003C\u002Fa>, \u003Ca href=\"\u002Ftag\u002Fopenclaw\">OpenClaw\u003C\u002Fa>, and Hermes, which is the right place to measure an \u003Ca href=\"\u002Ftag\u002Fagent\">agent\u003C\u002Fa> model if you care about real work. These tools differ in task shape, feedback style, and execution flow, so consistent performance across them suggests the model has learned transferable operating habits, not just benchmark-specific tricks. That is the difference between a model that demos well and a model that can be embedded into a team’s daily workflow.\u003C\u002Fp>\u003Ch2>The counter-argument\u003C\u002Fh2>\u003Cp>The skeptical view is straightforward: these are vendor-reported results, and the examples are curated. A 16-day autonomous coding run, a five-day research loop, or a quantified trading simulation can all be impressive while still hiding failure modes that matter in production. Long-horizon autonomy also creates a new risk surface: the model can drift, optimize the wrong objective, or appear competent while making subtle mistakes that only show up later. For many teams, a narrower model that is easier to supervise may still be the safer choice.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785844980448-n1d0.png\" alt=\"Qwen3.8-Max proves that agentic work is the real frontier\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>That objection is valid, but it does not defeat the release. It sets the standard for how to evaluate it. If a model claims agentic competence, then the right response is not to dismiss the demos, but to demand traces, reproducibility, and task-level audits. Qwen’s publication of long execution histories, commits, PRs, and benchmark runs is exactly the kind of evidence frontier agent models should provide. The limit is real: no model is trustworthy without domain-specific oversight. But the presence of that limit does not make the capability unimportant. It makes the capability operationally valuable only when paired with review.\u003C\u002Fp>\u003Ch2>What to do with this\u003C\u002Fh2>\u003Cp>If you are an engineer, PM, or founder, treat Qwen3.8-Max as a signal to redesign work around supervised autonomy. Use it where tasks have clear artifacts, testable outputs, and logs you can inspect: code generation, research pipelines, document review, multimodal extraction, and workflow automation. Do not buy the fantasy that the model replaces judgment. Buy the more useful reality that it can compress the middle of the work, leaving humans to set goals, review edge cases, and approve outcomes. The companies that win here will not be the ones asking for better chat. They will be the ones building better control loops around models that can already do the job.\u003C\u002Fp>","Qwen3.8-Max matters because broad, stable agentic work beats narrow benchmark wins.","zhuanlan.zhihu.com","https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F2067570518102905363",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785844979892-cdc1.png","model-release","en","8b6f5cee-1836-4999-b4ae-7e30d74c4799",[17,18,19,20,21,22],"Qwen3.8-Max","agentic work","real-world RL","long-horizon autonomy","Claude Code","OpenClaw",[24,25,26],"Qwen3.8-Max is best understood as an agentic work model, not a chat model.","Its strongest claim is breadth across coding, office work, research, and multimodal tasks.","The real test is supervised autonomy: traces, audits, and reproducible outputs.",1,"2026-08-04T12:02:32.157065+00:00","2026-08-04T12:02:32.148+00:00",{"tags":31,"relatedLang":34,"relatedPosts":38},[32],{"name":21,"slug":33},"claude-code",{"id":15,"slug":35,"title":36,"language":37},"qwen38-max-is-built-for-delivery-not-chat-zh","Qwen3.8-Max不是更会聊天，而是更会交付","zh",[39,45,51,57,63,69],{"id":40,"slug":41,"title":42,"cover_image":43,"image_url":43,"created_at":44,"category":13},"ccdf0c22-70d3-474f-9fa1-592694c564b1","google-earth-should-not-ship-ai-image-generation-en","Google Earth Should Not Ship AI Image Generation","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785805372500-cnzw.png","2026-08-04T01:02:30.317498+00:00",{"id":46,"slug":47,"title":48,"cover_image":49,"image_url":49,"created_at":50,"category":13},"6e9aa97d-d130-4c68-a2cd-d7bd78b7c610","try-claude-opus-4-7-benchmarks-safety-en","Try Claude Opus 4.7 and read its benchmarks","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785720761463-1tp9.png","2026-08-03T01:32:20.969744+00:00",{"id":52,"slug":53,"title":54,"cover_image":55,"image_url":55,"created_at":56,"category":13},"40da5e56-c978-4c19-b039-71d4121a46eb","opus-5-cut-cost-without-losing-quality-en","Opus 5 lets you cut cost without losing quality","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785607382907-76ym.png","2026-08-01T18:02:40.854976+00:00",{"id":58,"slug":59,"title":60,"cover_image":61,"image_url":61,"created_at":62,"category":13},"b3fd7185-d626-4e48-ae4d-d38170255e54","openai-cuts-gpt-56-prices-ai-bills-en","OpenAI Cuts GPT-5.6 Prices as AI Bills Climb","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785542568883-mf9l.png","2026-08-01T00:02:28.341618+00:00",{"id":64,"slug":65,"title":66,"cover_image":67,"image_url":67,"created_at":68,"category":13},"2fee41e9-10d2-4755-9777-081159b6e609","opus-5-premium-ai-becoming-commodity-en","Opus 5 proves premium AI is becoming a commodity","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785499372952-zhqg.png","2026-07-31T12:02:29.544844+00:00",{"id":70,"slug":71,"title":72,"cover_image":73,"image_url":73,"created_at":74,"category":13},"5ff2c0fd-3296-48d3-8c74-acdadf2d5605","openai-free-gpt56-access-scientists-en","OpenAI Gives Scientists Free GPT-5.6 Access","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785434595001-84cv.png","2026-07-30T18:02:38.762824+00:00",[76,81,86,91,96,101,106,111,116,121],{"id":77,"slug":78,"title":79,"created_at":80},"d4cffde7-9b50-4cc7-bb68-8bc9e3b15477","nvidia-rubin-ai-supercomputer-en","NVIDIA Unveils Rubin: A Leap in AI Supercomputing","2026-03-25T16:24:35.155565+00:00",{"id":82,"slug":83,"title":84,"created_at":85},"eab919b9-fbac-4048-89fc-afad6749ccef","google-gemini-ai-innovations-2026-en","Google's AI Leap with Gemini Innovations in 2026","2026-03-25T16:27:18.841838+00:00",{"id":87,"slug":88,"title":89,"created_at":90},"5f5cfc67-3384-4816-a8f6-19e44d90113d","gap-google-gemini-ai-checkout-en","Gap Teams Up with Google Gemini for AI-Driven Checkout","2026-03-25T16:27:46.483272+00:00",{"id":92,"slug":93,"title":94,"created_at":95},"f6d04567-47f6-49ec-804c-52e61ab91225","ai-model-release-wave-march-2026-en","Navigating the AI Model Release Wave of March 2026","2026-03-25T16:28:45.409716+00:00",{"id":97,"slug":98,"title":99,"created_at":100},"895c150c-569e-4fdf-939d-dade785c990e","small-language-models-transform-ai-en","Small Language Models: Llama 3.2 and Phi-3 Transform AI","2026-03-25T16:30:26.688313+00:00",{"id":102,"slug":103,"title":104,"created_at":105},"38eb1d26-d961-4fd3-ae12-9c4089680f5f","midjourney-v8-alpha-features-pricing-en","Midjourney V8 Alpha: A Deep Dive into Its Features and Pricing","2026-03-26T01:25:36.387587+00:00",{"id":107,"slug":108,"title":109,"created_at":110},"bf36bb9e-3444-4fb8-ab19-0df6bc9d8271","rag-2026-indispensable-ai-bridge-en","RAG in 2026: The Indispensable AI Bridge","2026-03-26T01:28:34.472046+00:00",{"id":112,"slug":113,"title":114,"created_at":115},"60881d6d-2310-44ef-b1fb-7f98e9dd2f0e","xiaomi-mimo-trio-agents-robots-voice-en","Xiaomi’s MiMo trio targets agents, robots, and voice","2026-03-28T03:05:08.899895+00:00",{"id":117,"slug":118,"title":119,"created_at":120},"f063d8d1-41d1-4de4-8ebc-6c40511b9369","xiaomi-mimo-v2-pro-1t-moe-agents-en","Xiaomi MiMo-V2-Pro: 1T MoE Model for Agents","2026-03-28T03:06:19.238032+00:00",{"id":122,"slug":123,"title":124,"created_at":125},"a1379e9a-6785-4ff5-9b0a-8cff55f8264f","cursor-composer-2-started-from-kimi-en","Cursor’s Composer 2 started from Kimi","2026-03-28T03:11:59.132398+00:00"]