Alibaba open-sources Qwen3.8-27B as 2.4-trillion Max weights debut

Alibaba released the weights of its 2.4-trillion-parameter Qwen3.8-Max model, the first open-source frontier-class model from a Chinese lab, challenging closed rivals Fable 5 and GPT-5.6 Sol on agent benchmarks. The Qwen3.8-2.4T-A95B, posted Aug. 13 on ModelScope and Hugging Face, activates 95 billion parameters per token and natively supports a 262,144-token context window, scalable to 1.01 million tokens.

"This may be the most powerful model besides Fable 5," the Qwen team said in its release announcement, framing the open-weights drop as a direct challenge to Anthropic's flagship. The smaller Qwen3.8-27B, a natively multimodal dense model with 27 billion parameters, went live Aug. 14 and outperforms the prior Qwen3.7-Plus, Alibaba said.

On PaperBench, a paper-reproduction benchmark, Qwen3.8-Max scored 93.0, ahead of GPT-5.6 Sol at 90.5, Fable 5 and Claude Opus 4.8. It also led OSworld-Verified with 86.1 and parametric CAD with 91.5. But the model trails on coding: SWE-bench Pro at 67.7 versus Fable 5's 80.0 and Claude Opus 4.8's 69.2, and TerminalBench 2.1 at 86.6 versus GPT-5.6 Sol's 88.8.

The release carries a commercial twist. Alibaba abandoned the permissive Apache 2.0 license used for earlier Qwen models in favor of a custom Qwen3.8-Max License that requires any MaaS or AI work-assistant business with more than $50 million in annual revenue to negotiate a separate commercial deal. Alibaba shares jumped 7 percent in Hong Kong and 4.5 percent on the NYSE after the Aug. 3 announcement of the monetization plan.

A License That Gates Scale

The $50 million revenue threshold replaces the era of free frontier-class weights with conditional access. The license defines MaaS broadly, covering any third-party access to inference or fine-tuning where the provider controls inputs or parameters. Smaller startups get a safe harbor, but any entity reaching significant scale — the ones most likely to threaten Alibaba's own API business — must negotiate terms. The model charges 12 yuan per million tokens for input and 36 yuan for output in mainland China, with overseas pricing at $2 and $6 per million tokens.

The strategy mirrors Moonshot AI's Kimi K3, which required commercial providers generating over $20 million in annual revenue to strike a deal with up to 30 percent revenue share. American cloud providers including DigitalOcean, Together AI and Fireworks AI have already signed commercial agreements around Chinese open-weight models under this emerging structure.

Benchmarks Need Independent Verification

Alibaba's vendor-reported figures rely on cross-harness comparisons that are not directly comparable. Qwen3.8-Max was evaluated on OpenCode, while Claude was tested on Claude Code and GPT-5.6 on Codex — environments that differ in timeout constraints, token limits and methodology. The Qwen team did not disclose test conditions for all comparisons.

The model's focus is long-horizon agent work. Alibaba ran Qwen3.8-Max through 16 days of continuous autonomous programming, had it reproduce and improve a research paper, and operate a virtual e-commerce enterprise across more than 2,000 rounds of interaction. The post-training process for office and agent scenarios combines continuous expansion of real-world environments, a unified reward system and an online data balancer.

The open-weights release strengthens Alibaba's cloud position as it competes with DeepSeek, Moonshot AI and xAI's Grok 4.6, all of which updated flagship models this week. The 27B dense variant is expected to drive most community adoption, while the restrictive license on the Max model gates the high-end enterprise market. Alibaba's cloud unit, which reported accelerating AI-related revenue growth, stands to capture both developer mindshare and enterprise licensing fees. The question for builders is whether the $50 million threshold forces consolidation around Alibaba's platform or pushes successful companies to migrate to alternative architectures.

This article is for informational purposes only and does not constitute investment advice.