Explore additional models
Frequently
asked questions
Qwen3.6 27B is a dense 27B-parameter open-weight variant of Qwen3.6. The Alibaba team released it in April 2026, following the February release of the Qwen3.5 series. It has a native 262,144-token context length and runs in thinking mode by default. It also adds thinking preservation, an option to retain reasoning context from historical messages. It ships under the Apache 2.0 license and is available for high-performance inference via Self-Serve Deployments in Crusoe Intelligence Foundry.
No. Qwen3.6 27B is a dense model that runs all 27B parameters on every token. Its mixture-of-experts sibling, Qwen3.6 35B A3B, activates about 3B of its 35B parameters per token, so it uses less compute per token. Both models use a hybrid attention design. Alibaba also trained Qwen3.6 27B for multi-token prediction, which serving frameworks can use for speculative decoding.
Alibaba focused this release on agentic coding, particularly frontend workflows and repository-level reasoning. Alibaba built Qwen3.6 on direct feedback from the community, with a stated focus on stability and real-world utility. Thinking preservation also suits iterative development, where each turn builds on the reasoning from earlier ones.
Yes. You can serve it with high-performance inference. Self-Serve Deployments run Qwen3.6 27B on your own dedicated GPUs with Crusoe’s optimized serving stack, so you don’t have to set up a vLLM or SGLang server or tune inference frameworks yourself. Choose an inference profile optimized for throughput or latency.
Thinking preservation is an option to retain reasoning context from historical messages. By default, Qwen3.6 27B keeps only the thinking content from the latest user message. With thinking preservation on, it keeps reasoning from earlier turns, so it can build on that reasoning instead of working through the same steps again. Alibaba trained Qwen3.6 to use these historical traces and recommends the option for agent scenarios and iterative development, where each turn depends on the last.


%204.png)
