알리바바가 장기적인 복잡 과제를 자율적으로 수행하는 2.4조 개 매개변수의 오픈 웨이트 언어 모델 Qwen3.8-Max를 공개했습니다. 이 모델은 16일간의 자율 소프트웨어 개발, 연구 논문 재현 및 성능 개선, 자동화된 칩 설계 등을 통해 서방 국가 최고 수준 모델들과 필적하는 성능을 입증했습니다. 모델 웨이트는 다음 주 공개될 예정입니다.
번역된 본문
알리바바의 새로운 플래그십 모델인 Qwen3.8-Max는 연구 논문을 재현하고 칩을 자율적으로 설계하는 등 수일에 걸쳐 복잡한 작업을 독자적으로 처리할 수 있도록 설계되었습니다. 팀은 다음 주에 웨이트(weights)를 공개할 계획입니다.
알리바바의 Qwen 팀은 현재까지 가장 성능이 뛰어난 언어 모델인 Qwen3.8-Max를 발표했습니다. 이 모델은 총 2.4조 개의 매개변수(parameters)로 확장되며, 쿼리당 950억 개가 활성화됩니다. Qwen3.8-Max는 Qwen3.5 아키텍처를 기반으로 구축되었으며, 팀은 단발성 프롬프트에 답하는 것보다 장시간 복잡한 작업을 독립적으로 완료하는 데 중점을 두었다고 밝혔습니다. 알리바바는 7월 중순에 표준 가격의 10%로 알리바바의 토큰 플랜(Token Plan), 코더(Qoder) 및 코더워크(QoderWork)를 통해 제공되는 프리뷰 버전으로 이 모델을 처음 발표했습니다. 당시에도 팀은 2.4조 개의 매개변수를 언급하며 모델을 Fable 5 다음으로 높게 평가했지만 벤치마크는 공유하지 않았습니다. Qwen3.8-Max는 웨이트(weights)가 공개될 첫 번째 Qwen-Max 급 모델입니다.
세 번의 자율 코딩 실행으로 모델의 능력 입증
Qwen3.8-Max의 코딩 능력을 선보이기 위해 팀은 모델이 인간의 도움 없이 작동한 세 가지 사례를 제시했습니다.
첫 번째 사례에서 Qwen3.8-Max는 커맨드라인 도구인 oh-my-cli를 구축하는 데 16일을 보냈습니다. 이 모델은 들어오는 사용자 요청을 받아 깃허브(GitHub) 이슈로 변환하고, 스스로 할당한 뒤 코드를 작성하고, 테스트를 실행하여 결과를 반복적으로 개선했습니다. 2026년 7월 30일까지 인간의 개입 없이 총 265개의 커밋(commits), 127개의 풀 리퀘스트(pull requests), 151개의 이슈를 처리했습니다.
두 번째 사례에서 모델은 연구 논문 "Unified Data Selection for LLM Reasoning"을 전달받았으나 시작 코드는 제공받지 않았습니다. 그 임무는 논문의 결과를 재현한 뒤 이를 개선하는 것이었습니다. 팀에 따르면 약 5일 동안 약 125시간의 컴퓨팅 시간 동안 Qwen3.8-Max는 7,600줄의 코드를 작성하고 33개의 GPU 학습 작업을 실행했습니다. 모델은 먼저 논문의 6가지 주요 결과를 모두 재현했습니다. 그런 다음 4라운드에 걸쳐 18가지 자체 아이디어를 테스트했으며, AIME24 수학 벤치마크에서 논문의 기존 방식을 2.7포인트 앞섰습니다.
세 번째 사례는 알리바바 톈치(Tianchi) 플랫폼에서 열린 WWW2025 멀티모달 대화 의도 인식 챌린지(526개의 인간 팀 참가)였습니다. 24시간 이내에 이 모델은 제품 스크린샷을 위해 Qwen2.5-VL-7B와 함께 여러 중국어 모델을 미세 조정(fine-tuned)하고 이를 투표 시스템으로 결합했습니다. 45번의 제출 결과 정확도가 0.60에서 0.853로 상승했습니다. 이 성적으로 Qwen3.8-Max는 526개 팀 중 458개 팀보다 앞섰습니다.
칩 설계 및 모의 회계 연도 테스트로 장기 계획 능력 평가
두 가지 추가 사례는 수백 번의 상호작용 라운드에 걸쳐 이루어지는 작업을 목표로 합니다. 첫 번째 사례에서 Qwen3.8-Max는 암호화 체계를 위한 암호학적 빌딩 블록을 설계해야 했습니다. 이러한 회로의 핵심 효율성 지표는 칩의 기본 요소인 논리 게이트(logic gates)의 수입니다. 게이트가 적을수록 더 작고 효율적인 칩을 의미합니다. 이 모델은 8,298개의 게이트를 사용하는 작동은 하지만 비대한 설계로 시작하여 약 500회의 반복을 거쳐 678개의 게이트로 줄였습니다.
오픈 소스 도구인 OpenROAD를 이용한 자동 레이아웃 작업 이후, 물리적 칩 면적은 106x106 마이크로미터에서 46x46 마이크로미터로 81% 감소했습니다. Qwen 팀은 수백 번의 반복 후에도 이 모델이 표면적인 수준에 머무르지 않고 계속해서 깊은 구조적 변경을 수행했다고 밝혔습니다.
Alibaba’s open-weight Qwen3.8-Max takes on long-horizon AI tasks with 2.4 trillion parameters Jonathan Kemper View the LinkedIn Profile of Jonathan Kemper Aug 3, 2026 Key Points Alibaba has unveiled Qwen3.8-Max, a 2.4-trillion-parameter language model designed to complete complex tasks on its own over multiple days. In testing, the model autonomously built software, reproduced and improved on a research paper's results, and successfully ran a simulated e-commerce business. Internal benchmarks put its performance on par with top Western models. Qwen3.8-Max is available now, with weights set to drop next week. Ask about this article… Search Alibaba's new flagship model Qwen3.8-Max is built to handle complex tasks on its own over days at a time, from reproducing research papers to designing chips autonomously. The team plans to release the weights next week. Alibaba's Qwen team has unveiled Qwen3.8-Max, its most capable language model to date. The model scales to 2.4 trillion total parameters, with 95 billion active per query. Qwen3.8-Max builds on the Qwen3.5 architecture, and the team says the focus is on completing complex tasks independently over extended periods rather than just answering one-off prompts. Alibaba first announced the model in mid-July as a preview version available through Alibaba's Token Plan, Qoder, and QoderWork at ten percent of the standard price. Even then, the team cited 2.4 trillion parameters and ranked the model just behind Fable 5 , but didn't share benchmarks. Qwen3.8-Max is the first model in the Qwen-Max class whose weights will be made publicly available. Ad Three autonomous coding runs put the model through its paces To show off Qwen3.8-Max's coding chops, the team presented three case studies in which the model worked without any human help. Ad DEC_D_Incontent-1 In the first, Qwen3.8-Max spent 16 days building the command-line tool oh-my-cli. The model took incoming user requests, turned them into GitHub issues, assigned them to itself, wrote the code, ran tests, and improved the results iteratively. By July 30, 2026, it had racked up 265 commits, 127 pull requests, and 151 issues, all without a single human touch. In the second case, the model received the research paper "Unified Data Selection for LLM Reasoning" but no starter code. Its job was to reproduce the paper's results and then improve on them. Over roughly five days and about 125 hours of compute time, Qwen3.8-Max wrote 7,600 lines of code and ran 33 GPU training jobs, according to the team. It first reproduced all six of the paper's main results. Then it tested 18 of its own ideas across four rounds and beat the paper's method on the AIME24 math benchmark by 2.7 points. Ad The third case involved the WWW2025 Multimodal Dialogue Intent Recognition Challenge on Alibaba's Tianchi platform, where 526 human teams competed. Within 24 hours, the model fine-tuned several Chinese language models along with Qwen2.5-VL-7B for product screenshots and combined them into a voting system. Across 45 submissions, accuracy climbed from 0.60 to 0.853. That put Qwen3.8-Max ahead of 458 of the 526 human teams. Chip design and a simulated fiscal year test long-horizon planning Two more case studies target tasks that stretch across hundreds of interaction rounds. In the first, Qwen3.8-Max had to design a cryptographic building block for encryption schemes. The key efficiency metric for such a circuit is the number of logic gates it needs, the basic elements on a chip. Fewer gates mean a smaller, more efficient chip. The model started with a working but bloated design using 8,298 gates and whittled it down to 678 gates over roughly 500 iterations. Ad DEC_D_Incontent-2 After an automated layout pass with the open-source tool OpenROAD, the physical chip area shrank from 106x106 to 46x46 micrometers, an 81 percent reduction. The Qwen team says the model kept making deep structural changes even after hundreds of iterations instead of settling for surface-level tweaks. Ad The second case study is E-Commerce-Bench, a simulation of an entire fiscal year in online retail based on anonymized data from Taobao and Tmall. The model starts with 100,000 yuan in capital and has to run multiple online stores in parallel for a full year. That means buying products, negotiating with suppliers in natural language, adjusting prices, managing returns, and dealing with crises like typhoons or supply chain disruptions. Hidden in the supplier pool are 152 scammers that the model has to spot. Qwen3.8-Max ended up with a balance of 416,252 yuan, quadrupling its starting capital. That's 38 percent more than the runner-up GLM 5.2 and more than 2.5 times what its predecessor Qwen3.7-Max managed. The model invested aggressively early in the year and pulled in a net profit of over 100,000 yuan during the holiday season. Multimodal skills and app reconstruction without source code For multimodal tasks, Qwen3.8-Max can process documents with over 200 pages and videos longer than 100 hours, the team says. They're also introducing RecreationBench, a new benchmark that requires the model to rebuild running applications without access to source code. The model can only observe the target app through interaction, meaning clicks and keyboard input. Testing covers Ubuntu, macOS, Windows, Android, and the web. Alongside this, the team is releasing Qwen-MM-Plugins, an extension library that adds image and video processing, visual tool use, and multimodal memory to existing agent systems. In the benchmark tables the Qwen team published, the model lands near or above Claude Opus 4.8, Claude Fable 5, and GPT-5.6 Sol across many categories. On PaperBench, Qwen3.8-Max hits 93, the highest score in the comparison. On TerminalBench 2.1, it scores 86.6, trailing GPT-5.6 Sol's 88.8. As is typical with self-reported numbers from model makers, these results come from internal runs. Independent verification is still pending. The Qwen team attributes the model's ability to sustain such long-running tasks to a major expansion of training environments during reinforcement learning. Training no longer focused on single tasks alone but also covered multi-day workflows, nested directory structures instead of individual files, and a variety of agent harnesses. The team's internal score index across more than ten benchmarks rose from 0.474 to 0.725. The model performed best at around 4,000 environments, after which scores dipped slightly. China's open-model race heats up Qwen3.8-Max's most direct rival also comes from China. Moonshot AI released Kimi K3 with open weights on Hugging Face on July 27 , a multimodal mixture-of-experts model with 2.8 trillion parameters and a one-million-token context window . Along with the weights, Moonshot also published parts of its own infrastructure, including attention kernels, an MoE communication library, and tools for running agents at scale. Independent testing tempered the company's claims, though. K3 fell well short of top Western models in both cyber capabilities and complex math . Qwen3.8-Max is available now through QwenCloud. The weights are set to go live on Hugging Face and ModelScope next week. The model supports both OpenAI's Chat Completions format and Anthropic's API protocol, so it plugs directly into Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw. A parameter called reasoning_effort lets users choose between three levels that trade speed for thoroughness. Qwen3.8-Max isn't the only piece Alibaba has added recently. A few weeks ago, the team introduced Qwen-Image-3.0, an image generator for information-dense layouts that handles inputs up to 4,500 tokens and renders readable text as small as ten pixels. At the other end of the scale, Alibaba continues to push small open models like Qwen3.6-35B-A3B , which has 35 billion total parameters but only activates three billion at a time. AI News Without the Hype – Curated by Humans Subscribe to THE DECODER for ad-free reading, a weekl