GLM-5.3 is now open-weight

The strongest practical caveat was speed and total cost in use, not raw capability. Multiple people said GLM-5.3-Flash can be slow or inconsistent depending on provider, and some reported that Chinese models still overthink, generating far more tokens than Anthropic or OpenAI for the same task. That can erase the apparent price advantage even when per-token rates look low. Provider choice mattered a lot. Several people said OpenRouter works well only if you pin a provider, otherwise routing changes kill prompt cache reuse and make costs jump. The local inference conversation was more sober than the hype. Almost nobody claimed self-hosting wins on pure economics. The dominant view was that cloud inference is still cheaper, faster, and less painful because memory is expensive, consumer hardware is in flux, and power and cooling costs are brutal. What kept local setups attractive was control. People want fixed behavior, no silent model swaps, no policy churn, no rate-limit games, no trust in zero data retention promises, and the ability to run offline or under stricter privacy and compliance constraints. Some also pointed out that local hardware can gain effective value over time as better open models arrive, even if the box never penciled out on a token-per-dollar basis. A second thread spun off into what open weights mean for model access and safety politics. The release was taken as more evidence that open models are catching up fast, and as an implicit rebuke to earlier claims that publishing strong models was too dangerous. The pushback was narrower and more concrete than the usual culture-war version. Older closed models are unlikely to be released not because they are still frontier-capable, but because they may leak training data, unsafe behaviors, or proprietary implementation details. On cyber capability, people mostly treated the debate as already decided. Fine-tuning for offensive use is already possible, so the practical priority now is defensive use at scale.