On August 3, Alibaba officially unveiled its next-generation foundation model, Qwen3.8, sporting an impressive total parameter count of 2.4 trillion. This new model marks a giant leap in coding and professional collaborative work (Cowork). On the authoritative LMSYS Chatbot Arena leaderboard, Qwen3.8 ranks in the global top tier, closely trailing Anthropic's Claude series. The API for its flagship version, Qwen3.8-Max, is now live on the Qianwen AI platform and integrated into Alibaba's new Agentic product "Qianwen Cowork," while Qwen3.8-27B and the Max version are slated for open-source release next week.
Pricing-wise, Qwen3.8 offers highly competitive rates. Domestically, the API costs 12 RMB per million tokens for input and 36 RMB for output, with implicit cache hits dropping to just 1.5 RMB. Globally, its input and output prices are only 40% and 24% of Claude 3.5 Opus, offering elite intelligence at an unbeatable performance-to-price ratio.
Featuring vision capabilities and a 1M-token context window, Qwen3.8-Max adopts an optimized Sparse #MoE and Hybrid Attention architecture with 95B active parameters. It scored a record-breaking 93.0 on the PaperBench coding agent benchmark and recorded 81.9 and 52.4 on WideSearch and Agent’s Last Exam, respectively. Crucially, it topped the OSWorld-Verified benchmark with 86.1, demonstrating superior computer-use capabilities for agents.
Autonomous coding is a core highlight of Qwen3.8, which ranks 4th globally on CodeArena. Unlike legacy models that only auto-complete code snippets, Qwen3.8 can build an entire project from scratch over several days without human intervention. By deploying a "Loop Engineering" framework, it successfully created a self-evolving agent framework called “oh-my-cli” (now open-sourced on GitHub) over a 16-day autonomous run.
In professional coworker scenarios, Qwen3.8 leverages joint reinforcement learning (RL) across real-world environments and compute power. It can label thousands of clauses in legal files in under an hour (a task that takes a legal team a week), slice 160 hours of sports video in minutes, or orchestrate parallel agents to design complex, multi-year ETF quantitative strategies. In long-horizon tasks like digital chip design or business simulations, Qwen3.8 showcases highly sophisticated "execute-feedback-iterate" closed-loop learning.
Qwen3.8 also pushes the boundaries of visual understanding, ranking 2nd globally on Vision Arena. In the source-free RecreationBench, it successfully reconstructed web applications solely through visual and interactive feedback. Additionally, Alibaba's Zhenwu M890 supernode has been fully optimized for Qwen3.8, delivering up to a 1.5x speedup in agentic reasoning workloads.
[AgentUpdate Depth Analysis] The debut of Alibaba’s Qwen3.8 represents a pivotal shift in LLM evolution—from simple text generation to complex, long-horizon autonomous agency. By marrying a colossal 2.4T MoE architecture with deep reinforcement learning, Qwen3.8-Max doesn't just undercut Western rivals like Claude 3.5 Opus on price; it effectively solves the "closed-loop execution" bottleneck on benchmarks like OSWorld and CodeArena. Its ability to self-correct and coordinate multi-agent systems over long periods (such as in hardware design or quantitative trading) pushes AI closer to L4 autonomy. As the 27B and Max open-source releases roll out next week, Qwen3.8 is poised to democratize state-of-the-art Agent development, challenging closed-source monopolies and triggering an explosion of practical, production-grade AI Agents globally.
