Published: April 21, 2026
Alibaba released Qwen3.6-Max-Preview on April 20, 2026, the most capable model in its Qwen series to date. The proprietary model achieves the highest scores on six major coding and agent benchmarks, including SWE-bench Pro and Terminal-Bench 2.0, while using a mixture-of-experts architecture that activates only 3 billion of its 35 billion total parameters per inference. Qwen3.6-Max-Preview is available now through Alibaba Cloud’s Bailian platform and Qwen Studio, with API compatibility for both OpenAI and Anthropic specifications.
What can Qwen3.6-Max-Preview do?
Qwen3.6-Max-Preview is a text-only large language model built for agentic programming tasks, complex reasoning, and instruction following. The model supports a 256,000-token context window and includes a preserve_thinking feature that carries reasoning traces across multi-turn conversations, a capability Alibaba specifically recommends for agentic workflows where continuity of context is essential.
The model uses a mixture-of-experts (MoE) architecture with 35 billion total parameters but activates only 3 billion parameters per inference call. This design significantly lowers computing costs and increases throughput compared to dense models of equivalent capability. At launch, Qwen3.6-Max-Preview does not accept image inputs; it is a text-in, text-out model.
Qwen3.6-Max-Preview benchmarks and technical specifications
Alibaba reports that Qwen3.6-Max-Preview ranks first on six coding and agent benchmarks: SWE-bench Pro (real-world software engineering), Terminal-Bench 2.0 (command-line execution), SkillsBench (general problem-solving), QwenClawBench (tool use), QwenWebBench (web interaction), and SciCode (scientific programming). These benchmarks cover a broad range of practical developer tasks, from fixing real GitHub issues to orchestrating multi-step tool calls.
Compared to its predecessor Qwen3.6-Plus, the Max-Preview variant shows gains of +9.9 points on SkillsBench, +10.8 on SciCode, and +3.8 on Terminal-Bench 2.0. World knowledge scores also improved, with +2.3 points on SuperGPQA and +5.3 on QwenChineseBench. Instruction-following capability, measured on ToolcallFormatIFBench, improved by +2.8 points.
How does Qwen3.6-Max-Preview compare to GPT-5.4 and Claude Opus 4.7?
The April 2026 AI model landscape is highly competitive. On software engineering tasks,
Claude Opus 4.6 leads SWE-bench Verified at 80.8%,
while Qwen3.6-Plus (the predecessor to Max-Preview) scored 78.8% on the same benchmark.
GPT-5.4 scored 57.7% on SWE-bench Pro.
With the additional gains reported for Qwen3.6-Max-Preview over Qwen3.6-Plus, Alibaba’s latest model appears to be closing the gap with the leading proprietary models on coding-specific tasks.
A notable distinction is the pricing and architecture. Qwen3.6-Max-Preview activates only 3 billion parameters per request despite having 35 billion total, making it substantially more efficient to run than dense models like GPT-5.4 (priced at $2.50/$15 per million tokens) or Claude Opus 4.7 ($5/$25 per million tokens). Alibaba has not yet announced final pricing for Qwen3.6-Max, though the predecessor Qwen3.6-Plus is currently free during its preview period. The model’s context window of 256,000 tokens is smaller than the 1 million+ tokens offered by GPT-5.4 and Claude Opus 4.7.
Qwen3.6-Max-Preview availability and access
Qwen3.6-Max-Preview is available immediately through Alibaba Cloud’s Bailian platform and Qwen Studio. The API endpoint uses the model string qwen3.6-max-preview and is compatible with both the OpenAI and Anthropic API specifications, allowing developers to integrate it into existing pipelines with minimal code changes.
Alibaba explicitly labels this release as a preview. The company states that the model is “still in active development” and that further optimizations are expected before a stable release. No open weights are available; the model is hosted and proprietary.
What does the Qwen3.6-Max-Preview release mean for the AI market?
The release of Qwen3.6-Max-Preview adds another strong contender to the April 2026 wave of AI model launches. Alongside GPT-6 (April 14), Claude Opus 4.7 (April 16), Google Gemma 4 (April 2), and Meta Llama 4 (April 5), it confirms that Chinese AI labs are shipping models that compete directly with Western frontier systems on technical benchmarks. The efficient MoE architecture, activating just 3 billion of 35 billion parameters, offers a different trade-off than the dense models from OpenAI and Anthropic, prioritizing inference cost over raw parameter count.
Full details on Qwen3.6-Max-Preview are available on the Qwen research page and through the Alibaba Cloud Model Studio.
