Came away from this amazed at how much latent ability models possess. DeepSeek-V4-Pro solves 51% of SWE-bench Pro with one attempt and 71% with 64 attempts. Efficient inference turns parallel sampling into a lever worth pulling, especially on tasks you can verify.
1/ We think parallel test-time scaling is one of the most underexplored ways to improve model qualityView quoted post on X
We spend a lot of time on efficiency for long-horizon agentic inference
We showed that a swarm of DeepSeek-V4-Pro agents lifts the SWE-Bench-Pro score by ~20 points, beating the cost-performance trade-off of the leading frontier models🧵