Eight of the ten most-used models on OpenRouter’s latest weekly LLM leaderboard were built in China. Collectively, those models handled roughly 29.55 trillion tokens — about 87% of the top-ten traffic on the platform’s weekly table. On Monday, Alibaba added another layer of pressure to that trend. The company unveiled Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model that Alibaba claims ranks “second only to Fable 5” among frontier systems. The release is significant for two reasons. First, it is Alibaba’s largest Max-class model to date. Second, it marks the first time Alibaba plans to open the weights of a Max-class flagship. Those weights are scheduled to land on Hugging Face and ModelScope during the week of August 10. As of publication, no repository, license, or model card had appeared on either platform. The timing is not accidental. Less than a month before this launch, Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model that currently holds the record for the largest open-parameter system. So Qwen3.8-Max is not the largest open-weight model on paper. But it may be the most strategically important release from Alibaba in years.
Qwen3.8-Max is built on the Qwen 3.5 architecture. It uses a sparse mixture-of-experts design that activates only 95 billion parameters during inference — roughly 4% of its total parameter count. Alibaba says this keeps computational costs and latency closer to a 100-billion-parameter model than to a 2.4-trillion one. The model also supports a context window of up to 1 million tokens, enough for documents beyond 200 pages, video footage beyond 100 hours, or roughly 750,000 words of input. It natively handles text, image, and video inputs. | Specification | Qwen3.8-Max | |---|---| | Total parameters | 2.4 trillion | | Active parameters (MoE) | 95 billion | | Architecture | Sparse mixture-of-experts | | Context window | 1 million tokens | | Multimodal support | Text, image, video | | Open-weight release | Week of August 10, 2026 | | Distribution platforms | Hugging Face, ModelScope | The message from Alibaba is clear: it wants a seat at the frontier table, not just in the open-source community.
Benchmark wins — and a few losses
Alibaba published a broad set of benchmark numbers alongside the launch. On some tests, Qwen3.8-Max outruns both Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol. On others, it trails. | Benchmark | Qwen3.8-Max | Fable 5 | GPT-5.6 Sol | |---|---:|---:|---:| | PaperBench | 93.0 | 88.8 | 90.5 | | OSWorld-Verified | 86.1 | 85.0 | 83.2 | | SWE-bench Pro | 67.7 | 80.0 | — | | IFBench | 82.8 | 63.5 | — | | TerminalBench 2.1 | 86.6 | — | 88.8 | | VideoMME v2 | 68.3 | — | 71.1 | The strongest results are in document reproduction, computer-use agents, and general instruction-following. Qwen3.8-Max beats Fable 5 on PaperBench and OSWorld-Verified, and it leads by a wide margin on IFBench. The less flattering numbers are on software engineering. On SWE-bench Pro, Qwen3.8-Max scores 67.7, trailing Fable 5’s 80.0 and Claude Opus 4.8’s 69.2. On TerminalBench 2.1, it falls slightly behind GPT-5.6 Sol. On long-video understanding, it also trails. There is an important caveat. None of those GA-era numbers have been independently reproduced by a neutral testing platform yet. The benchmarks are Alibaba’s own disclosures, and the community has not had enough time with the production model to verify them.
The 16-day coding experiment
The most striking claim in Alibaba’s announcement is not a score. It is a story. Alibaba says Qwen3.8-Max independently completed a software engineering project over 16 days without human intervention. The model reportedly accumulated 265 commits, 127 pull requests, and 151 issues in a self-built GitHub repository for a project called oh-my-cli. In another demonstration, the model competed against 526 human teams in the WWW2025 multimodal dialogue intent recognition challenge on Alibaba’s Tianchi platform. According to Alibaba, it finished ahead of 87% of the human field within 24 hours. The company also claims the model reproduced and improved a research paper on AI reasoning over roughly 125 hours, gaining 2.7 points on the AIME24 benchmark through self-evolution. Those are impressive claims. They are also unverified. “The claim worth examining is not the parameter count. Alibaba says the model completed a software engineering project in 16 days. That sentence has been reprinted everywhere and interrogated nowhere,” said Amit Jena, development manager for AI at Kanerika. “Sixteen days of what? How many times did a human step in? Did the output survive code review?” That skepticism is not unfair. Long-running autonomous coding tasks are notoriously difficult to reproduce, and Alibaba’s demonstrations remain company-produced. Still, they point to a broader shift in the industry: frontier models are increasingly judged by their ability to finish workflows, not just answer prompts.
Pricing that undercuts the West
Qwen3.8-Max is already available through Alibaba’s QwenCloud APIs. The pricing structure is startlingly aggressive. | Region | Input (per 1M tokens) | Output (per 1M tokens) | Cache hit | |---|---:|---:|---:| | China | ¥12 (~$1.70) | ¥36 (~$5.10) | ¥1.50 (~$0.21) | | International | $2.00 | $6.00 | $0.25 | At $6 per million output tokens, Qwen3.8-Max costs about 40% of Kimi K3’s $15, roughly one-fifth of GPT-5.6 Sol’s $30, and slightly more than one-ninth of Claude Fable 5’s $50. Alibaba claims these prices make Qwen3.8-Max the most cost-efficient option in its performance class. The market responded quickly. Its U.S.-listed shares also gained around 4.5% in premarket trading.
[SPONSORED]
▶ ENTERPRISE GPU CLUSTERS ◀
Scale your AI model training seamlessly. Book a Demo.
The community is paying attention
On Hacker News, the Qwen3.8-Max thread passed 684 points within hours of publication. The reaction was not uniform. “From my experience Qwen-3.7-Max is above the Opus level but delivers results much faster. Slightly worse than Fable. Way ahead of Deepseek 4 Pro — which is a workhorse on its own,” one user wrote. Coding4Food, which tracks Hacker News sentiment, noted a cluster of early developers calling the model “surprisingly smooth.” Not all preview feedback was positive. During the preview phase, a Linux DO user claimed Alibaba had initially served the wrong checkpoint — suspected to be Qwen3.7 Max Preview — which reportedly failed classic visual reasoning tests. The claim was never confirmed externally. A separate Reddit post during the preview phase also claimed a jailbreak prompt could bypass the model’s safety filters. That post remains unverified as well. Despite the noise, developer activity around the model has been fast. Multiple pull requests have already been merged into Qwen repositories to support reasoning effort parameters, image modality, and prompt cache optimizations.
Open-weight does not yet mean open-license
Alibaba’s product page describes Qwen3.8-Max as open-source. But as of publication, there was no repository on Hugging Face or ModelScope, no license, and no exact release date beyond “week of August 10.” That gap matters. “Publishing weights is a separate act from opening an API endpoint,” Jena said. “Until there is a repository, a licence and a model card, open-weight describes an intention.” Alibaba’s licensing history does not make the next step obvious. Qwen2.5-72B shipped under a custom Qwen commercial license. Qwen2.5-3B was research-only. Most other Qwen2.5 models were released under Apache 2.0. Which regime Qwen3.8-Max will fall under is still unknown. If Alibaba follows the restrictive path, enterprise adoption will slow. If it chooses a permissive license, Qwen3.8-Max could become the most downloaded open-weight model of the year.
The larger enterprise shift
Forrester analyst Charlie Dai sees the launch as evidence that open-weight models have crossed a credibility threshold. “Enterprises increasingly have credible alternatives to proprietary frontier models, particularly for software engineering, domain customization, sovereignty, and cost-sensitive deployments, where openness often matters as much as absolute model performance,” Dai said. This is not just a China narrative anymore. Gartner predicts that the share of global enterprises including Chinese LLMs and multimodal models in their AI portfolios will rise from 5% in 2025 to 50% by 2027. The Financial Times has reported that a large share of U.S. AI startups now use Chinese open-source models, with companies such as DoorDash, Siemens, and Airbnb deploying them in production. Microsoft, according to The Information, is also testing Moonshot’s Kimi K3 for some Copilot inference workloads. The report estimated that switching certain workloads away from OpenAI and Anthropic models could save Microsoft as much as $600 million per year in cloud infrastructure costs. Alibaba Cloud does not need to wait for the mega-enterprise deals to feel the momentum. In the company’s 2026 fiscal fourth quarter, AI-related product revenue crossed 30% of total cloud segment revenue for the first time, with annualized revenue above 35.8 billion yuan. MaaS token revenue grew 15x over the five months ending in May.
What happens next
The next week will be more revealing than the last one. A weight release without a usable license will not move enterprise procurement decisions. A permissive license, by contrast, could push Qwen3.8-Max into production environments faster than most proprietary rivals. eWeek’s coverage of the launch put the challenge plainly: Alibaba needs to prove that Qwen3.8-Max can move beyond demos to reliable production use. A 2.4-trillion-parameter MoE model still requires serious infrastructure to run at scale, and long-running autonomous tasks remain prone to failure from bad reasoning, unexpected errors, or poor decisions. The benchmark tables will generate clicks. The coding demo will generate debate. But the real test is the download page, the license, and whether independent teams can reproduce what Alibaba claims. Until then, Qwen3.8-Max is best understood as Alibaba’s strongest argument that the open-weight frontier no longer belongs to startups or Western labs alone. The weights — and the fine print — will decide how much that argument is worth.