alt Hacker News> They are also buying 300-400 employees with proven record of training good coding models.
Are they? Their Composer 2.5 models is based on Kimi K2.5, it's not a bespoke model.
> They are also buying 300-400 employees with proven record of training good coding models.
Are they? Their Composer 2.5 models is based on Kimi K2.5, it's not a bespoke model.
Achieving the improvements from K2.5 -> composer 2.5 with just post-training is actually more impressive. Though I believe their next model is trained from scratch.