Open-Weights Models & Serving Engines: the practice test
Running the 2026 open-weights frontier: GLM-5.3, Kimi K3 and DeepSeek V4. Reading config.json to size a model you have never run, latent attention and sparse indexers, vLLM and SGLang configuration, expert parallelism and all-to-all backends, weight formats, and the benchmarks that do not lie. This test drills exactly that: 11 easy, 14 medium and 10 hard questions, every one explained, every explanation linking into the worked material.
Sample questions, answered
Total parameters set the memory footprint because every expert must be resident somewhere. Active parameters bound decode speed because only the selected experts are read per token. The attention design sets KV per token, which now varies by more than an order of magnitude between models of similar size. And the released format decides bytes per parameter, which changes the first number by two to four times.
Those fields mean latent attention, so the cache holds one compressed vector plus the rotary part per layer rather than keys and values per head. That is (512 + 64) times 2 bytes times 78 layers, which is 89,856 bytes. Computing it from head count and head dimension instead gives about 3.66 MB, an overestimate of roughly 41.7 times, and that mistake turns a workable plan into an absurd one.
Go deeper than the quiz
A practice test measures recall. The material it draws from teaches the reasoning: