Tag: Model Serving

For the initial wave of post-training reasoning models, test-time compute was confined behind proprietary black-box APIs. Frontier systems withheld raw chain-of-thought tokens, masking internal reasoning trajectories behind encrypted tokens or high-level summaries. Builders were left with opaque cost structures, inability to fine-tune intermediate reasoning steps, and strict operational lock-in. The release of DeepSeek-R1 upended that […]