Meituan Open-Sources 1.6-Trillion-Parameter MoE Model LongCat-2.0 Under MIT License
Releasing a 1.6T parameter MoE model under the permissive MIT license is a massive escalation in the open-weight LLM ecosystem. For engineers, this removes the commercial use restrictions and revenue caps typically seen with Meta or Mistral models, opening the door for unrestricted enterprise-scale fine-tuning. The primary engineering challenge now shifts entirely to multi-node serving and inference optimization for a model of this unprecedented footprint.
On July 6, 2026, Chinese tech giant Meituan open-sourced LongCat-2.0, a massive 1.6-trillion-parameter Mixture of Experts (MoE) model. Most notably, the release falls under the highly permissive MIT License, marking a significant milestone in the scale of truly open-source AI models available for unrestricted commercial use.
Technical Details While the headline number is 1.6 trillion total parameters, the MoE architecture means the active parameter count during inference will be significantly lower, likely in the 100B-300B range depending on the specific routing mechanism (e.g., top-2 or top-4 expert routing). Despite the sparse activation, the VRAM footprint required to simply host the model weights will still demand a massive multi-node GPU cluster. Engineers will need to rely heavily on advanced quantization techniques (like FP8 or INT4), pipeline parallelism, and tensor parallelism to make serving LongCat-2.0 economically viable.
Why It Matters The choice of the MIT License is the most disruptive element of this release. Previous models approaching this scale have either been locked behind proprietary APIs or released under custom, semi-permissive licenses with commercial user limits or acceptable use clauses (such as Meta's Llama series). By utilizing the MIT license, Meituan has completely removed legal friction for enterprise adoption. Companies can now fine-tune, distill, or directly integrate a frontier-class MoE model into proprietary products without fear of licensing compliance issues, revenue caps, or vendor lock-in.
What to Watch Next The immediate bottleneck for the open-source community will be infrastructure. Watch for rapid updates from inference frameworks like vLLM, TGI, and TensorRT-LLM as they optimize for LongCat-2.0's specific MoE routing topology. Additionally, expect a wave of specialized, fine-tuned variants of LongCat-2.0, alongside smaller, distilled models that attempt to capture its performance at a fraction of the parameter count. Finally, independent benchmark evaluations will be critical to determine if Meituan's 1.6T scale actually translates to state-of-the-art reasoning capabilities against current proprietary leaders.