What happens when a model that can run locally starts matching Claude Sonnet on task reliability? Developers who've tested Qwen3.6-397B-A17B - Alibaba's latest model - are pushing Alibaba to release the weights as open source, and the case they're making is harder to dismiss than most community wishlist requests.
The model name explains the architecture: 397 billion total parameters, but only 17 billion are activated for any given response. This is a Mixture of Experts (MoE) design - instead of running the full network for every output, the model routes requests through small clusters of specialized "expert" sub-networks. The result is inference speeds (how fast the model generates a response) and compute costs closer to a 17B model, with capability that reportedly beats several larger competitors.
The Reliability Report
Benchmark scores aren't the main talking point. The consistent feedback from developers who've tested Qwen3.6-397B-A17B is about reliability - specifically, completing multi-step tasks from start to finish without losing context or producing half-finished results. Testers report it outperforms GLM-5.1 and Kimi-k2.5 on this dimension, and describe it as the first locally-runnable model that feels comparable to Claude Sonnet for autonomous work.
That framing is significant because reliability is exactly where local models have traditionally fallen apart. A model that scores well on coding benchmarks but abandons the task on step 7 of 10 isn't production-ready. The comparison to Sonnet specifically on end-to-end task completion - not just benchmark scores - suggests a real capability shift in this generation.
Why Alibaba Might Stay Closed
The commercial argument for keeping this model proprietary is straightforward: a model this capable has real API revenue potential. But Alibaba's history with the Qwen series cuts the other way. Qwen2.5 and its variants built one of the largest fine-tuning (customizing a model on specific data for a particular task) communities in open-source AI, producing hundreds of specialized variants that extended the original model's reach far beyond what Alibaba could have built alone.
A closed Qwen3.6 trades that community momentum for API revenue. Given the pattern with previous Qwen releases, the weights will likely come eventually - Alibaba has consistently favored open releases. The argument for moving quickly is that models this capable become less useful to independent developers the longer they sit behind an API paywall.