On 30 June 2026 Meituan released its new-generation foundation model LongCat-2.0 and launched an open-source programme in parallel. According to public information, LongCat-2.0 uses a Mixture-of-Experts architecture with 1.6 trillion total parameters and native pre-training from scratch, and is the first trillion-scale model to complete the full pipeline from pre-training to inference entirely on domestic compute.
Why “full-pipeline domestic compute” matters
Training and running a trillion-parameter model demands extreme scale and stability in computing power. In the past such R&D relied heavily on imported compute. The significance of LongCat-2.0 is that it makes “running both training and inference entirely on domestic compute” a precedent.
From an industry perspective, this means domestic compute has moved from “able to support small and medium models” to “able to support ultra-large-scale models” — a node in the process of compute self-reliance. It also echoes the policy emphasis on “ultra-large-scale intelligent computing clusters” and “strengthening high-quality data supply”.
Open source: a technical route and a compliance route
LongCat-2.0 launched an open-source programme in parallel. Open source is an accelerator for the model ecosystem — lowering the barrier to use, widening application and attracting co-development. But from a compliance angle, open source also means a whole new set of responsibility allocations.
First, the choice and understanding of open-source licences. Different licences impose different limits on commercial use, modification and redistribution. Companies using open-source models must read the licence terms and avoid mistaking “open source” for “unconditional free use”.
Second, downstream transmission of model capability responsibility. Upstream open source does not mean downstream immunity. Applications built on an open-source model must still perform their own filing, registration and content-labelling duties.
Third, the tension between training data and open source. Distributing an open-source model may carry out latent rights defects in training data, and the model provider must assess the data-compliance exposure that open source creates.
What it means for companies using open-source models
First, filing and registration duties are not waived because “an open-source model is used”. Applications or features that call on already-filed model capabilities through APIs or otherwise must complete registration; those that fine-tune and provide services externally must assess whether filing is triggered.
Second, build an open-source component inventory. Centrally registering the licences of the models, frameworks and tools used is a basic action for compliance review and IP risk.
Third, note the ownership of outputs. The IP ownership of fine-tuning, distillation and optimisation based on an open-source model must be determined by the licence and contracts.
Compliance support along the domestic-compute chain
As model R&D moves from “buying compute” to “using domestic compute across the full pipeline”, companies also face new contractual and compliance arrangements: compute procurement and scheduling agreements, the data boundary between compute providers and model developers, the sourcing and authorisation of training data, and the compliance of cross-border compute use. These should be clear early in a project.
For support on open-source model compliance, filing and registration, and compute cooperation contracts, see our Algorithm Filing and Commercial Contracts services.
This article is based on public information on the release of LongCat-2.0 on 30 June 2026. It is for reference only and does not constitute legal advice.