Introducing SWE-2: Pushing the Pareto Frontier | Cognition Post-trained from a 2.8T-parameter model using Pareto-informed cost penalties in RL, the new coding agent matches top benchmarks on FrontierC

Introducing SWE-2: Pushing the Pareto Frontier | Cognition Post-trained from a 2.8T-parameter model using Pareto-informed cost penalties in RL, the new coding agent matches top benchmarks on FrontierCode 1.1 while cutting costs by up to 64%. It is rolling out immediately across Devin's desktop, CLI, and web tools. https:// cognition.com/blog/swe-2

1 reportother

Introducing SWE-2: Pushing the Pareto Frontier | Cognition Post-trained from a 2.8T-parameter model using Pareto-informed cost penalties in RL, the new coding agent matches top benchmarks on FrontierC

mastodon:mstdn-socialother17d ago kagi ↗

Introducing SWE-2: Pushing the Pareto Frontier | Cognition Post-trained from a 2.8T-parameter model using Pareto-informed cost penalties in RL, the new coding agent matches top benchmarks on FrontierC

Claim audit

All coverage

Introducing SWE-2: Pushing the Pareto Frontier | Cognition Post-trained from a 2.8T-parameter model using Pareto-informed cost penalties in RL, the new coding agent matches top benchmarks on FrontierC