Contrastive Language Model is 9x faster than Jev, sets SOTA on agentic coding benchmarks

iamrobotbear · x · 2026-09-24

A new System One model, Contrastive Language Model (CLM), contrasts with Jev's RLCD-trained typed-decision approach: CLM embeds states and candidate actions and ranks them by similarity. CLM-8B, pretrained on internet-scale data, offers up to 9x faster inference than Jev and, with light fine-tuning, sets SOTA on DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%), while Jev reportedly fails as a verifier on long-horizon tasks. The author urges developers building custom harnesses to watch this wave and links a guide on combining System One and System Two models.

Related event: CLM-8B: Contrastive Language Model Tops Agentic Coding Benchmarks at 9x Speed(3 posts)→

Original post →

More from coding & agent

coding & agent channel →