CLM: An Open-Weights Alternative to TypeSafe AI's Jev Decision Model

Contrastive Language Model (CLM) has been released as an open-source, self-hostable projection head for Qwen models (specifically CLM-v0.1-8B), offering full functional parity with TypeSafe AI's proprietary Jev decision model. It implements the exact same decision primitives—Choice, Noul, and Score—allowing developers to use it as a drop-in replacement for the Jev client API. CLM allows developers to replace proprietary cloud-based decision APIs with lightweight local models (~75 MB heads), delivering 4x to 13x latency speedups through disaggregated action caching. Furthermore, it enables open fine-tuning on custom agent trajectories, outperforming Jev on coding verifier benchmarks such as Terminal-Bench 2.1 and DeepSWE. While CLM achieves feature parity and higher benchmark scores after fine-tuning, Jev retains an advantage in zero-shot broad knowledge tasks and supports up to a 64K token context compared to CLM's calibrated 2K–8K range. Additionally, CLM computes candidate probabilities relatively using dot products and softmax, whereas Jev relies on absolute internal calibration.