On April 20, Alibaba released Qwen3.6-Max-Preview, the most powerful model in the Qwen series to date.
Notably, this launch came without a grand open-source celebration, no Apache license announcement, and no banners welcoming free downloads or commercial use. Qwen3.6-Max is only accessible via Qwen Studio and Alibaba Cloud Model Studio API; the weights are not publicly available.
This marks a turning point in Alibaba's AI strategy, and it deserves a closer look.
Benchmark Results: Six First-Place Finishes, Second Overall
First, the numbers. Qwen3.6-Max achieved first place in six mainstream benchmarks:
| Benchmark | What It Tests |
|---|---|
| SWE-bench Pro | Real-world software engineering tasks |
| Terminal-Bench 2.0 | Command-line execution ability |
| SkillsBench | General problem-solving |
| QwenClawBench | Tool calling |
| QwenWebBench | Web interaction |
| SciCode | Scientific programming |
After independent evaluation by Artificial Analysis, the model was ranked second overall—behind only Meta's Muse Spark and well above the median of other reasoning models in its price range.
Compared to its predecessor, Qwen3.6-Plus, the Max version improved by 2.3% on SuperGPQA (high-difficulty comprehensive reasoning) and by 5.3% on the Chinese language test QwenChineseBench. These are genuine gains, not the result of test-swapping or score-chasing.
Technical Specifications
Context window: 256k tokens. Notably, this is much smaller than Qwen3.6-Plus's 1 million tokens—Alibaba made a trade-off here, opting for a shorter window in exchange for higher reasoning quality.
Input format: Currently text-only, with no image support. This is somewhat unexpected for a flagship model, but the Preview version is expected to add it later.
API compatibility: Compatible with both OpenAI and Anthropic API specifications, making switching costs very low.
Preserve_thinking: A noteworthy new feature—the model can carry its reasoning trace across multiple conversation turns. It remembers how it reasoned before, rather than starting from scratch each time. This is useful for scenarios requiring long, multi-step interactions.
Key Question: Why Did the Open-Source Giant Suddenly Close Its Flagship?
The Qwen series previously followed a typical open-source path. Qwen3 used the Apache license, and Qwen3.6-35B-A3B was open-sourced as recently as April 16, with weights released.
But Max is different this time.
Alibaba is not the first to take this route:
- Meta built a developer ecosystem with Llama's open-source approach, then launched Muse Spark as its first closed-source flagship from a super-intelligence lab.
- Now Alibaba has drawn a line at its strongest point.
The logic behind this is not hard to understand. Open source is good for capturing market share and building an ecosystem, getting developers on your side. But the top-tier models are the real assets for landing big clients and collecting service fees. Open-sourcing small models while charging for large ones—this has become the mainstream strategy in 2026.
In fact, the entire industry is converging in this direction: OpenAI's closed-source flagship, Meta's closed-source Muse Spark, Alibaba's closed-source Max. The only ones still fully open are mainly small-to-medium research labs and projects with explicit ideological commitments.
What This Means for Developers
Good news: The API is compatible with OpenAI and Anthropic interfaces. If you are already using one of them, switching requires almost no code changes.
Bad news: Local deployment is not possible. For scenarios with high data privacy requirements, low-latency priorities, or offline operation, Max is not an option. In such cases, continue using Qwen3.6-35B-A3B.
Neutral news: 256k context is sufficient for most practical tasks. A 1-million-token window is impressive, but few scenarios actually use it to the full. Max concentrates resources on reasoning quality, a reasonable trade-off.
What's Next
Currently, Max-Preview is in preview status. Based on past practice:
- Image support will be added in the official release.
- Pricing has not been announced yet; there is no pricing during the public beta.
- The open-source small model ecosystem (35B-A3B, etc.) will continue to be maintained and will not be discontinued.
Alibaba's AI strategy is shifting from broad ecosystem building to open-source for attracting users and closed-source for monetization. In the 2026 AI market, this is almost an inevitable choice—because running purely open source at a certain scale leaves a very narrow path to commercialization.
For developers who chose Qwen because of its open-source nature, this shift is worth noting. The flagship is now closed, but the smaller models remain. How you use them depends on your scenario.
Sources: CocoLoop, Alibaba Drops Qwen 3.6 Max Preview—Its Most Powerful Model Yet (Decrypt)