A Bigger Model Won't Save You From a Vague Spec

Meituan open-sourced LongCat-2.0 last week: 1.6 trillion parameters, a native 1-million-token context window, and SWE-bench Pro numbers that edge out GPT-5.5. It trained entirely on Chinese ASICs, ships under MIT with weights on Hugging Face, and has been quietly topping OpenRouter. It is a genuinely impressive piece of engineering. It will not fix the thing that is actually breaking your codebase.

The industry keeps treating agentic coding as a capability problem. Bigger model, longer context, higher benchmark, better agent. So every few weeks a new frontier drops and teams re-run the same experiment: point it at a real repo, watch it produce plausible code, watch that code drift from what anyone actually asked for. The model got smarter. The output got worse in the ways that matter.

Here is the part nobody selling you a model wants to say out loud. Model capability is not your bottleneck. Intent is.

Consider the 1M-token context window everyone is excited about. The pitch is that the agent can now hold the whole codebase in its head, so it stops forgetting your constraints. In practice, a million tokens of undifferentiated context is a million tokens the model can weight however it likes. If your requirements live in a Slack thread, three stale comments, and one engineer's memory, a bigger window just means the agent has more room to confidently guess wrong. Context is not specification. Volume is not intent.

This is why spec-driven development keeps winning the argument even as the models underneath it churn. A versioned, executable spec is the one artifact that does not get better or worse when the model changes. It is the fixed point. Swap GPT-5.5 for LongCat-2.0 tomorrow and your spec still says the same thing, your acceptance criteria still fail the same way, your review still catches the same drift.

Three things worth doing this week, ranked by payoff:

1. Write the spec before you pick the model. If your intent only exists as a prompt, you have coupled your correctness to whatever model happened to be behind the endpoint that day. Decouple them. The spec is the contract; the model is an interchangeable implementer.

2. Treat benchmark wins as noise until they are independent. Every LongCat-2.0 number so far comes from Meituan's own testing. A sub-one-point SWE-bench Pro lead over GPT-5.5, self-reported, is not a reason to migrate anything. It is a reason to keep your evaluation harness pointed at your own tasks.

3. Make drift a review gate, not a vibe. If you cannot diff the generated code against a stated intent, you are not reviewing, you are proofreading. The spec is what turns "does this match what we asked for" into a checkable question instead of a feeling.

Open-weight models this strong are great news. Cheaper inference, no vendor lock, real competition at the frontier. But a 1.6-trillion-parameter model is still a very fast way to build the wrong thing quickly if nobody wrote down what the right thing was. The model was never the hard part. Deciding what you actually want, precisely enough that a machine can be held to it, always was.