DeepSeek published a training recipe that claims frontier-grade reasoning with a fraction of the GPU-hours used by comparable RL runs, sparking immediate debate among labs optimizing post-training stacks.
The method in plain terms
The method combines curriculum-style difficulty ramping, selective replay of failure modes, and a smaller critic model that gates expensive rollouts — only promising trajectories receive full GPU budget.
On public math and code suites cited in the paper, results land within a few points of much larger baselines, though independent replication is still early. Several teams note sensitivity to base-model choice and tokenizer quirks.
Only promising trajectories get full GPU budget — a gated rollout loop that trims waste.
Replication watch
For practitioners, the practical takeaway is cost routing: cheaper RL loops may let smaller orgs iterate on reasoning behaviors without reserving entire clusters for weeks.
DeepSeek released partial logs and ablation configs under a research license; weights remain on their existing open-weight terms for the accompanying checkpoint.
