One-sentence summary
Using Meituan's fully domestic-trained LongCat-2.5, evaluate domestic-chip-domestic-model claims by training scale, end-to-end domestic coverage, performance and ecosystem, not just params.
Why evaluate separately
LLM training demands extreme compute-stack stability, interconnect and software tooling; whether domestic compute can run trillion-parameter models is its own question, not just a benchmark score.
Steps
- Check training scale: LongCat-2.5 at 1.6T tests domestic compute throughput and stability.
- Check end-to-end coverage: training and inference fully on domestic compute matters more than partial use.
- Check real performance: usability on long context, multimodal and agent tasks.
- Check ecosystem: whether compilers, scheduling and fault-tolerance tooling are ready.
Common Mistake
Stopping at it ran on domestic chips, ignoring end-to-end coverage and real performance, overstates or understates domestic maturity.
Sources
- Meituan: Meituan launches LongCat-2.5