Back to feed
News Story
APriority79
AI前线
1 sources

Alibaba Releases Qwen3.8: 2.4T Parameters, Enhanced Coding Abilities

Alibaba has officially released its next-generation foundation model Qwen3.8, with a total of 2.4 trillion parameters, showing significant improvements in coding and professional work capabilities. The model performs exceptionally well on authoritative benchmarks, and its API is now available on the Qwen AI platform, alongside the launch of the enterprise agent product 'QwenWork'. Qwen3.8-Max is expected to be open-sourced next week, along with Qwen3.8-27B.

SynthePulse Insight · AI deep reading

Alibaba Releases Qwen3.8 with 2.4 Trillion Parameters: 16 Days of Autonomous Programming, Significant Gains in Coding and Office Capabilities

Version 1 · 1 source

On August 3, Alibaba officially released its new generation foundation model Qwen3.8, with a total of 2.4 trillion parameters, achieving breakthroughs in coding and office capabilities, and simultaneously launching the enterprise-grade Agent product 'QwenWork'.

  • Qwen3.8 has a total of 2.4T parameters, with 95B activated, using sparse MoE and hybrid attention mechanisms, and a context length of 1M tokens.
  • In the PaperBench coding evaluation, Qwen3.8-Max improved by 28.2 points over the previous generation, setting a new high of 93.0 points.
  • Qwen3.8 autonomously programmed for about 16 days, building the self-evolving agent framework 'oh-my-cli' from scratch, which has been open-sourced.
  • 'QwenWork' has started public beta, with web and PC clients available, and DingTalk integration coming soon.
  • Qwen3.8-Max is expected to be open-sourced next week, along with Qwen3.8-27B.
Open section navigationModel Release and Core Specifications

Model Release and Core Specifications

On August 3, Alibaba officially released its new generation foundation model Qwen3.8, with a total of 2.4 trillion parameters, making it the largest and most powerful flagship model in the Qwen series. The model employs a joint optimization of sparse MoE architecture and hybrid attention mechanisms, with 95B activated parameters, a context length of 1M tokens, and support for visual understanding.

The API for Qwen3.8 is now available on the Qwen AI platform, priced at 12 yuan per million tokens for input and 36 yuan for output domestically, with implicit cache hits at only 1.5 yuan; internationally, prices are 40% and 24% of Opus 5, respectively. Additionally, Alibaba's Zhenwu M890 supernode has been adapted for Qwen3.8, enabling up to 1.5x performance improvement in Agentic reasoning scenarios.

Breakthroughs in Coding and Office Capabilities

In coding capabilities, Qwen3.8 ranks fourth globally on the authoritative CodeArena leaderboard. According to official statements, Qwen3.8 can autonomously complete real project deliveries lasting over ten days, starting from an empty folder, without any human intervention. For example, with just a single instruction to 'create a self-evolving agent harness,' Qwen3.8 autonomously built a loop engineering framework, independently programmed for about 16 days, and produced a self-evolving agent framework 'oh-my-cli' at the level of Hermes Agent, which has been fully open-sourced.

In office capabilities, Qwen3.8 achieved real performance improvements across hundreds of high-value, high-frequency professional tasks through joint reinforcement learning expansion in real environments and computing power. For instance, a legal assistant team would need a week to annotate over a thousand clauses, but Qwen3.8 completed it within an hour; a basketball data analysis team would need to annotate 160 hours of video frame by frame, but Qwen3.8 broke down over 8,400 offensive and defensive rounds in tens of minutes; a financial research team would need years of accumulated quantitative strategy research, but Qwen3.8 delivered a complete ETF rotation strategy after several hours of continuous work.

Evaluation Results and Visual Capabilities

In multiple evaluations, Qwen3.8-Max leads: PaperBench coding agent evaluation scored 93.0, up 28.2 points from the previous generation; WideSearch and Agent's Last Exam general agent evaluations scored 81.9 and 52.4, respectively; IF Bench instruction following scored 82.8; GPQA Diamond scientific reasoning scored 92.6; BabyVision visual reasoning scored 82.0, nearly double that of some mainstream models; OSWorld-Verified computer operation scored 86.1, ranking first among mainstream models.

In visual capabilities, Qwen3.8-Max ranks second globally on the Vision Arena leaderboard. It can read 200-page financial report PDFs, understand long videos exceeding 100 hours, and feed visual information back into the entire workflow. In the 'app replication' benchmark RecreationBench, the model replicated an entire application from scratch through interaction and feedback alone, without source code or internet access.

Agent Product and Open Source Plans

Alibaba also launched the enterprise-grade Agent product 'QwenWork' (Qianwen Office), starting public beta. Individual and enterprise users can experience it via the official website (qwenwork.cn), with web and standalone PC clients currently available, and DingTalk PC and mobile built-in entries opening soon. Users can experience the Qwen3.8 flagship model within it.

Regarding open source, Qwen3.8-Max is expected to be open-sourced next week, along with Qwen3.8-27B. Starting today, developers worldwide can access Qwen3.8 API services through the Qwen AI platform.

Credibility boundary

The information in this article is primarily sourced from AI Front's coverage of Alibaba's launch event, which is a second-hand account. Specific evaluation scores and performance data are official claims and have not been independently verified. Some data, such as 'Hermes Agent level,' are official descriptions and should be treated with caution.

Insight takeaway

Qwen3.8 demonstrates Alibaba's technical progress in foundation models with its 2.4T parameters and autonomous programming capabilities, but actual performance still requires verification by third-party evaluations and the open-source community.

Primary report

AI前线

Primary source