Contrastive-LM released CLM-8B, an open-weights System One model designed to score candidate agent actions instead of generating text. Released under an Apache-2.0 license, the 75 MB head utilizes a frozen Qwen3-8B encoder served via vLLM and targets the API interface of TypeSafe AI’s proprietary Jev model.
CLM-8B trains state and action encoders using a bidirectional InfoNCE loss, scoring candidates via embedding dot products. In agent loops where states change constantly but action sets remain static, CLM-8B caches action vectors to reduce state-revisit latency from 1.7 ms to 0.6 ms on an RTX 4090 GPU, running up to 13 times faster than Jev with 1,000 candidates.
The research team reports new state-of-the-art verifier results on DeepSWE and Terminal-Bench 2.1 benchmarks, running 4.1× to 5.7× faster than baseline verifiers on an H100 GPU.
Why it matters
Replaces generative LLM text outputs with efficient vector-based action scoring, drastically cutting agent latency and inference costs.
Provides an open-source, Apache-2.0 alternative to proprietary System One agent controllers, enabling self-hosted agent routing.
Source: marktechpost.com



