Composer 2
Notes on Composer 2's training stack, covering its Kimi K2.5 base, two-stage continued pretraining and RL pipeline, distributed rollout infrastructure, and NVFP4 GEMM kernels that make long-running coding-agent training practical.
Notes on Composer 2's training stack, covering its Kimi K2.5 base, two-stage continued pretraining and RL pipeline, distributed rollout infrastructure, and NVFP4 GEMM kernels that make long-running coding-agent training practical.