On July 26, Meituan’s team open-sourced LongCat-2.0, a model with 1.6 trillion total parameters but only 48 billion active per inference. The same week, the KwaiKAT team at Kuaishou released KAT-Coder-V2.5, a 35-billion-parameter MoE model trained inside real executable repositories. And in the background, reporters from Reuters and a detailed analysis by Ella Apostoaie painted a picture of Beijing wrestling with whether to keep its best models open or lock them down.
These three stories touch the same nerve. The open-source coding model space is moving fast, and the Chinese government is starting to wonder if that speed is safe.
LongCat-2.0 is big, but not in the way that matters most
Meituan’s LongCat-2.0 has 1.6 trillion parameters. That number sounds impressive until you read the fine print: only 48 billion are active at any time. This is a Mixture-of-Experts architecture, and the 1.6T/48B ratio is the key design decision. Meituan claims the model is optimized for what they call Agentic Coding, meaning the model can act as an autonomous agent that writes, tests, and fixes code across multiple steps.
The architectural innovations are LongCat Sparse Attention and N-gram Embedding. Sparse attention matters for long-context handling. Code tasks often involve hundreds of lines of context, a function definition at the top and a bug at line 300. Standard transformers pay O(n^2) attention cost. Sparse attention cuts that. N-gram Embedding is a way to represent token sequences more efficiently, likely helping with code idioms that span multiple tokens like for i in range.
The more interesting part is that Meituan released inference code optimized for domestic Chinese computing cards. That is a practical move. Most Chinese AI labs depend on NVIDIA hardware, but U.S. export controls restrict access to the latest chips. By supporting domestic GPUs, Meituan makes LongCat-2.0 usable for local enterprises that cannot buy H100s. I do not know how well those domestic cards perform in practice, but the existence of optimized inference code is better than nothing.
KAT-Coder-V2.5 trains on real repositories, not synthetic snippets
The KwaiKAT team took a different approach. KAT-Coder-V2.5 is a 35B-total, 3B-active MoE model built on top of Qwen3.6-35B-A3B. They trained it on 100,000 verifiable repository environments. That means the model learns by actually running code in sandboxes, not by predicting the next token on a corpus of GitHub README files.
This is important. Most coding models are evaluated on static benchmarks like HumanEval, where the model generates a single function and a test checks it. Real coding is not single-turn. You write code, run it, see an error, fix it, run again. KAT-Coder-V2.5 was trained to operate inside that loop. The results show on PinchBench it scores 94.9, and on SWE-Bench Pro it scores 65.2, second only to Opus 4.8. Those numbers come from the closed flagship. The open-weight variant, KAT-Coder-V2.5-Dev, has separate benchmarks and is not directly comparable.
The open-weight version is under Apache-2.0. That matters for the broader discussion about openness.
China wants open models, but also wants control
This is where the third story fits. Apostoaie’s analysis covers a genuine tension. On July 16, Moonshot AI released Kimi K3, a 2.8 trillion parameter open-weight model that briefly topped a frontend coding leaderboard. The next day, President Xi Jinping told the World AI Conference to encourage open-source. Nine days earlier, Reuters reported that China’s Commerce Ministry had met with Alibaba, ByteDance, and others to discuss restricting access to advanced models.
The Chinese government has long promoted open-source software as a way to reduce dependence on foreign technology. Open-weight releases also shift inference costs to whoever deploys the model, which helps labs that cannot afford to serve millions of users. Moonshot AI already paused new subscriptions for Kimi K3 because demand pushed against capacity limits.
But frontier models have risks. Open weights cannot be recalled. Safety alignment can be removed with a few seconds of fine-tuning, as Princeton researchers showed with Llama 2. The U.S. Center for AI Standards and Innovation found that DeepSeek answered 94 percent of malicious requests under a common jailbreak, compared to 8 percent for U.S. references. Chinese regulators have started talking about circuit breakers and safety stop switches for frontier models.
Apostoaie argues that Beijing will move toward selective openness, keeping good-enough models open while restricting the ones that cross a dangerous capability threshold. The problem is that China currently lacks the evaluation infrastructure to draw that line. There is no formal government institution that tests for dangerous capabilities like CBRN risks or autonomous cyberattacks. The Chinese AI Safety and Development Association says it intends to develop evaluation methods, but has not done so yet.
What this means for builders
If you are building tools on top of these open models, the immediate takeaway is that you have two solid options for agentic coding. LongCat-2.0 gives you scale with efficient activation, and it runs on domestic hardware. KAT-Coder-V2.5 gives you a smaller, faster model trained on real execution environments. Both are open-weight under permissive licenses. You can download them and fine-tune them for your specific codebase.
The longer-term takeaway is that this openness may not last. If Beijing starts restricting frontier model releases, the models available today could become the ceiling for open access. That does not mean the models will disappear. It means the next generation of more capable models might be closed or gated. If your workflow depends on open-weight models, now is the time to evaluate and integrate, because the options next year may be more limited.
I do not know whether selective openness will work. The line between safe and dangerous is hard to draw, and models improve faster than safety evaluations. But the trend is clear. Chinese labs are producing some of the most capable coding models in the world, and they are releasing them openly. The government is watching, and it is starting to worry.