LLM Model Notes
Reading an open-weight model card
Separate the artifact from the announcement: revision, total and active parameters, native context, and the license file itself.
Independent reference
LLM Model Notes
Clear, practical guides to open-weight models from Chinese labs, evaluation design, inference constraints, and the trade-offs behind model comparisons.
Updated · See the project log
Current notes
Serving mixture-of-experts models
Why memory follows total parameters while per-token compute follows active ones, and what that means for hardware plans.
Measuring tokenizer efficiency for Chinese text
A reproducible way to compare how many tokens a fixed Chinese corpus costs under different tokenizers.
Reading benchmark results carefully
Define the evaluation contract, inspect uncertainty, and check prompt sensitivity before treating a score as evidence.
Latency, throughput, and context
Measure first-token delay, generation cadence, and capacity under a workload another person can reproduce.
Repeatable model comparisons
Turn a product question into a pinned manifest, paired observations, and a compact result bundle.