BREAKING
4 Frontier Models Tie on KV-Cache Math
One Prompt, One Hard Spec
1
Exact formulas
↓
2
5 presets
↓
3
Testing API
↓
4
11 checks
0
GiB
KV-cache size
0
x
attention reduction
0
%
work cached
Where the Builds Diverged
GPT-5.6 Sol
●
Strictly rejects out-of-spec inputs
Fable 5 & Grok 4.5
●
Relaxed API for testing
●
UI stays strict
GLM 5.2
flawed preset
●
Missed head-dim check
●
Silent defect
0
layers used
0
spec layers
0
x
inflation
Correctness vs Graceful Failure
AI NEWS BLITZ
Four frontier models nailed the KV-cache math, but split on how they validate the edges.