“First of many blogs” from Arcee.
AFM-4.5B scaled from 4K → 64K context.
⮕ arcee.ai/blog/extending…
𝑱𝑼𝑺𝑻 𝑴𝑬𝑹𝑮𝑬, 𝑫𝑰𝑺𝑻𝑰𝑳𝑳, 𝑹𝑬𝑷𝑬𝑨𝑻.
Proof it scales:
Same merge–distill cycle applied to GLM-4-32B.
Fixes 8K degradation in the 0414 release. +5% overall, strong