Alibaba Ships Qwen3.8-Max, Claims It Rivals Anthropic and OpenAI

Alibaba released its largest AI model yet on Monday, and it says the new system rivals the top models from Anthropic and OpenAI. The model, Qwen3.8-Max, is already live for developers on Vercel's AI Gateway — days after Alibaba first made it broadly available.

What Alibaba is claiming

Alibaba calls Qwen3.8-Max its "largest and most capable AI model to date," claiming performance that rivals the best systems from Anthropic and OpenAI as well as domestic rival Moonshot AI's Kimi K3. The company had previewed the model last month, when it said the system was "second only to Fable 5," Anthropic's flagship. Alibaba's own benchmark testing, shared Monday, backs that framing: the model's third-party ranking on the crowdsourced Arena.AI text leaderboard puts Qwen3.8-Max behind only Fable 5 and three models in Anthropic's Opus family. On frontend coding specifically, only two Claude Opus models and Kimi K3 rank above it; for visual analysis, only Fable 5 beats it.

The numbers behind the model

Alibaba says Qwen3.8-Max has 2.4 trillion parameters — the adjustable numerical settings a model learns during training and uses to process data, recognize patterns, and carry out tasks. That's a smaller count than Moonshot's Kimi K3, which has 2.8 trillion parameters, though parameter count alone is a rough proxy for capability at best: neither OpenAI nor Anthropic discloses exact parameter counts for their own top systems, so direct comparison across labs is only ever partial. Separately, Vercel's AI Gateway changelog lists the model with a context window of up to 1 million tokens, and says it's built to handle text and vision-language work in a single model — including turning screenshots or design files into working pages, captioning video, and answering questions grounded in an image.

Open weights, again

Alibaba says it will release the weights for Qwen3.8-Max next week. Weights are the adjustable values that determine how a trained model processes information; releasing them lets developers run, modify, and fine-tune the model themselves rather than only calling it through an API. The move marks a return to open-weight releases for Alibaba, after the company had pivoted toward keeping its more advanced models proprietary earlier this year. Open-weight releases have become the norm rather than the exception among Chinese AI labs — Moonshot released Kimi K3's weights just last week, and Beijing has openly framed the strategy as a way to grow China's influence over global AI governance and drive adoption of domestic AI systems abroad.

Where developers can already use it

Access isn't waiting on the weights drop. Vercel added Qwen3.8-Max to its AI Gateway on August 2, a day before Alibaba's own blog post went up, letting developers call the model by setting model to alibaba/qwen3.8-max in the AI SDK. Vercel's gateway reflects the provider's own pricing with no markup, including on bring-your-own-key requests, and developers can also point existing coding agents — Claude Code, Codex, OpenCode, or Pi — at the model by running the gateway's coding-agent setup command and selecting it inside the agent. That gives teams a low-friction way to benchmark Qwen3.8-Max against whatever model they're already running, without switching infrastructure providers first. Vercel also tracks real-world adoption independent of any benchmark: its AI Gateway model leaderboard ranks models by total token volume processed across all Gateway traffic, which will show whether developers actually route production work to Qwen3.8-Max or treat it as a benchmark curiosity.

The bigger competitive picture

Qwen3.8-Max's release lands in the middle of a rapid-fire stretch of Chinese model launches. It closely follows Kimi K3, and both ByteDance and MiniMax shipped new video-generation models the same week. Alibaba's release adds to already high tensions in Washington and Silicon Valley over how to manage the safety and export implications of increasingly capable open-weight AI systems out of China, especially as those systems narrow the performance gap with US frontier labs on independent leaderboards rather than just in vendor-published benchmarks.

That pace of Chinese open-weight releases has also revived a policy fight in the US. Amid reports of a potential crackdown on open models following this wave of Chinese launches, the US AI industry has largely rallied around preserving access to open-weight systems — arguing it's both a safety necessity and a way of preserving competition, rather than a reason to restrict access further. For teams evaluating Qwen3.8-Max, that policy backdrop matters as much as the benchmark numbers: an open-weight model you can inspect and self-host carries different long-term risk than a closed one you can only call through an API, regardless of which one scores higher this week.

What to watch next

Two things will tell you whether Alibaba's claim holds up outside its own testing: how Qwen3.8-Max's Arena.AI ranking moves once more independent voters weigh in, and whether real coding-agent usage — via Vercel's gateway or elsewhere — matches the frontend-coding benchmark Alibaba is citing. The open-weight release next week will also let independent researchers run their own evals rather than relying on Alibaba's numbers alone.