Contrastive-LM Releases CLM-8B Open-Source Agent Model

Contrastive-LM has released CLM-8B, an open-source single-model system designed to evaluate agent actions up to 9 times faster than Jev. Available under an Apache-2.0 license, the head weighs 75 MB and operates on a single NVIDIA GPU under Linux, utilizing vLLM to serve the Qwen3-8B encoder.

Jev previously entered limited early access on September 15, 2026, returning typed values with probabilities instead of raw text. CLM targets this same interface, providing a GitHub repository that serves CLM-8B behind a TypeSafe-compatible API and exposes three question types for users to replay requests through its Python client.

For its training methodology, CLM trains a state encoder and an action encoder using a bidirectional InfoNCE loss. Each encoder consists of a frozen Qwen3-8B backbone combined with a 20M-parameter trainable projection head. Training pulls each state toward its actual action while pushing it away from others, using dot products between state and action embeddings for scoring during inference.

This architectural design disaggregates states and actions. In an agent loop, the state changes at every step while the action set remains mostly static. Consequently, clm-serve reserves a slab of GPU memory similar to vLLM's KV cache to reuse cached vectors. On a single RTX 4090 with three actions, revisited states drop from 1.7 ms to 0.6 ms, with model cards reporting speeds up to 13x faster than Jev with roughly 1,000 candidates.

Regarding accuracy evaluation, pre-training alone achieves 52.1% top-1 accuracy across approximately 100K held-out questions, while mid-training raises it to 69.2%. The 9x figure originates from the T-Rex game, where actions repeat across states, with CLM matching Jev on T-Rex and Super Mario while trailing on tool calling and WikiRacing while maintaining superior speed across all tasks.