Alibaba has released Qwen3.8-Max. It is the newest and largest model in the Qwen family. The model builds on the Qwen 3.5 foundation. Alibaba plans to release its weights to the public. This marks a return to open releases after a stretch of keeping top models closed.
The launch lands in a busy week for Chinese AI labs. Moonshot AI released Kimi K3 last month. Both companies are racing to close the gap with US labs.
Qwen3.8-Max is live now through Alibaba Cloud’s Model Studio API. It also runs on QwenWork, Alibaba’s new workplace AI agent platform, which entered public beta the same day. Open weights follow about a week after this initial release.
Scale and Architecture
The model has 2.4 trillion total parameters. That makes it the second-largest model out of China. Only Moonshot’s Kimi K3, at 2.8 trillion parameters, is bigger.
Total parameter count only tells part of the story here. Qwen3.8-Max uses a sparse Mixture-of-Experts design. It pairs this with a hybrid attention mechanism. The model activates only about 95 billion parameters per query. This keeps costs and response times down compared to a dense model of the same size.
The model reads a context window of up to one million tokens. That equals roughly 750,000 words in a single query. Few models anywhere match this range.
Benchmark Performance
Alibaba shared its own benchmark numbers, and they place Qwen3.8-Max near the top of the field:
- Text Arena: 5th place globally
- Vision Arena: 2nd place globally, behind only a Claude Fable 5 variant
- Frontend Code Arena: 4th place, with a score of 1,668
Alibaba claims results equal to or better than Anthropic’s Fable 5 in several areas. These include multimodal reasoning, visual agent tasks, coding, and office work. Alibaba also points to scores close to OpenAI’s GPT5.6-Sol on some tests. Outside reporting tells a more careful story. Reviewers note strong results on select coding, multimodal, and engineering tests, and weaker results on general reasoning tests. Vendor benchmarks deserve a skeptical read. Companies pick the tests that show their models in the best light. Still, the pattern shows up across several outlets, and that points to a real jump for the Qwen line, not just marketing.
Multimodal and Agentic Capabilities
Alibaba wants Qwen3.8-Max to stand out beyond benchmark scores. The company markets it as a true multimodal tool. It reads hundred-page documents, full television series, or up to 100 hours of livestream footage. It turns all of this into a searchable knowledge base.
Alibaba highlights several demos on the creative and engineering side:
- Turning raw personal footage into polished video
- Making educational animations straight from a text prompt
- Rebuilding a front-end web project from one screenshot
- Turning a 2D floor plan into a 3D interior view
- Building a playable game from a plain-language description
Alibaba built a new benchmark to test these reconstruction claims. It calls this test RecreationBench. The model must rebuild an app from scratch in a closed environment. It gets no internet access and no source code. It works from interaction and visual feedback alone.
The boldest claim centers on autonomous coding. Alibaba ran an internal test. The model spent 16 days writing, testing, fixing, and refining an AI coding tool on its own, with little human input. If this result holds up in real use and not just a curated demo, it signals a real push into long-running, agentic software work. Anthropic and OpenAI are chasing the same goal.
Pricing and Availability
Alibaba priced Qwen3.8-Max to compete hard on cost. Input tokens run about 40 percent of the Claude Opus 5 rate. Output tokens run about 24 percent of that rate. Cache hits bring the price down further. Add the efficiency gains from the MoE design, and Qwen3.8-Max lands as one of the cheaper frontier-class options. This matters most for high-volume or long-context work.
How It Stacks Up
| Qwen3.8-Max | Kimi K3 (Moonshot) | Claude Fable 5 (Anthropic) | |
|---|---|---|---|
| Total parameters | 2.4T | 2.8T | Not disclosed |
| Active parameters | ~95B (MoE) | Not disclosed | Not disclosed |
| Context window | 1M tokens | 1M tokens | Not disclosed |
| Openness | Open weights (pending) | Open weights | Closed |
| Relative pricing | Below Opus/Fable tier | — | Baseline |
Qwen3.8-Max does not top the charts outright. It sits behind several Anthropic models on general text and vision tasks. It closes the gap fast, though, and it does so at a much lower price. It also gives developers a real path to self-host once the weights land.
Strengths
- A one-million-token context window fits document-heavy and long-running work
- The MoE design keeps inference costs low despite the huge parameter count
- Strong multimodal and reconstruction skills, backed by a dedicated benchmark
- Sharp, developer-friendly pricing next to comparable Western models
- An open-weight release plan, unlike several recent flagship models from other labs
Weaknesses
- It trails leading Anthropic and OpenAI models on general reasoning and text-arena rank
- The 16-day autonomous coding claim comes from an internal Alibaba test, not outside review
- Open weights are not out yet, so early users are stuck with the hosted API
- Some benchmark comparisons come straight from Alibaba, so weigh them against independent tests as they arrive
Verdict
Qwen3.8-Max does not knock the current frontier leaders off their perch. It proves the gap is shrinking fast, and it does this at a fraction of the cost. Its scale, efficient MoE design, huge context window, and bold multimodal demos make it one of the strongest open-weight-bound models to date. It is not the smartest model on the market by general benchmark scores. For teams that need a large context window, strong multimodal document and video reading, and low-cost pricing, with a future option to self-host, Qwen3.8-Max earns a real look.
