Alibaba Releases Qwen 3.8-Max, a 2.4 Trillion Parameter Model With Open Weights Planned
By OFOKI TECH | August 11, 2026
Alibaba’s Qwen team released Qwen 3.8-Max on August 3, 2026, positioning it as the most capable model in the Qwen family to date. The model features 2.4 trillion total parameters with approximately 95 billion active parameters per forward pass, using a sparse Mixture-of-Experts architecture. It is multimodal, supporting text, image, and video inputs, and offers a context window of up to 1 million tokens. Alibaba has stated that open weights for both Qwen 3.8-Max and a smaller Qwen 3.8-27B variant will be released, though as of August 10, the weights have not yet appeared on Hugging Face or ModelScope.
Architecture and Scale
Qwen 3.8-Max is built on a sparse MoE architecture, which means only a fraction of the 2.4 trillion parameters are activated for any given token. This design keeps inference costs manageable despite the massive parameter count. The model supports a context window of 1 million tokens, which translates to roughly 750,000 words per query, and can generate outputs up to 128,000 tokens long.
The model is available through Alibaba Cloud’s Model Studio APIs and the QwenWork productivity platform, which entered public beta on the same day. The API supports both OpenAI-compatible and Anthropic-compatible protocols, allowing developers to substitute Qwen 3.8-Max into existing toolchains with minimal configuration changes.
Benchmark Performance
Alibaba published a comprehensive benchmark package with the launch, though independent verification remains limited. On OSWorld-Verified, which measures how well an agent operates a real desktop environment, Qwen 3.8-Max scored 86.1, ahead of Claude Fable 5 at 85.0 and GPT-5.6 Sol Max at 83.2. On PaperBench, which tests a model’s ability to reproduce research paper experiments in code, it scored 93.0, the highest reported score, ahead of GPT-5.6 Sol at 90.5 and Fable 5 at 88.8.
On Terminal-Bench 2.1, a test of terminal-based engineering tasks, Qwen 3.8-Max scored 86.6, placing it between GPT-5.6 Sol at 88.8 and both Claude Opus 4.8 and Fable 5 at 84.6. On GPQA Diamond, a graduate-level science reasoning benchmark, it scored 92.6, level with Fable 5 and just behind GPT-5.6 Sol at 94.1.
The model shows weaker performance on SWE-bench Pro, which measures resolving real GitHub issues in professional codebases. There it scored 67.7, trailing Fable 5 at 80.0 and Opus 4.8 at 69.2, though ahead of GPT-5.6 Sol at 64.6.
Autonomous Coding Demonstrations
Alibaba highlighted two long-horizon demonstrations. In one, Qwen 3.8-Max reportedly coded autonomously for over 16 days, building a complete CLI tool called OMI CLI from an empty repository. The project trace, including over 500 commits, is public on GitHub. In another demonstration, the model participated in a simulated year-long e-commerce benchmark where it quadrupled its starting capital by managing multiple online stores, negotiating with suppliers, and identifying scam vendors.
These demonstrations are vendor showcases, not independent reproductions. The model also reportedly ran autonomously for 10 days building a production application from an empty folder, though the same caveat applies.
Pricing and Accessibility
Qwen 3.8-Max is priced at $2 per million input tokens and $6 per million output tokens, with cached input at $0.25 per million tokens. This undercuts several proprietary frontier models. A smaller Qwen 3.8-27B is also planned for open-weight release, targeting developers who want to run a capable model on consumer hardware. One community estimate suggests the 27B variant could run on approximately 17GB of RAM.
The open-weight release would mark the first time Alibaba has released weights for a Max-class model. Previous Max-tier releases, including Qwen 3.7-Max, remained API-only. The license terms for the open weights have not been disclosed, leaving open the question of whether they will use a permissive license like Apache 2.0 or a more restrictive custom license.
Market Position
Qwen 3.8-Max enters a crowded field of large language models released in mid-2026. It follows the open-weight release of Moonshot AI’s Kimi K3 at 2.8 trillion parameters, and joins other recent releases including DeepSeek V4 and GLM 5.2. On the LM Arena leaderboard, Qwen 3.8-Max ranks fourth overall, making it the second open-weight model to break into the top five.
Alibaba’s Hong Kong-listed shares rose 7% following the release, closing at HK$125.20 on August 3. The company has positioned QwenWork as a direct competitor to workplace AI platforms from other domestic and international providers.
Sources and further reading