- Published on
Contrastive-LM Releases CLM-8B: An Open System One Model That Scores Agent Actions Up to 9× Faster Than Jev
- Authors
- Name
- DeepHorus
Contrastive-LM has released CLM-8B, an open model designed to choose among proposed actions rather than generate prose. Instead of predicting a next token, it compares candidate actions with a representation of the current state and returns a probability distribution. The company positions it as the first openly available “Contrastive Language Model,” targeting agentic systems that need fast, typed decisions.
That approach follows a new interpretation of “System One” models. Proprietary Jev from TypeSafe AI entered limited early access on September 15, 2026, and exposes typed values with probabilities instead of conversational text. CLM closely matches that interaction contract: requests written for TypeSafe’s API can be replayed through its Python client, easing migration for developers.
Deployment is relatively lightweight. The Apache-2.0-licensed scoring head weighs 75 MB, while the surrounding system uses frozen Qwen3-8B encoders served through vLLM on Linux with a single NVIDIA GPU. CLM offers three question types: Noul for estimating statement truth, Choice for ranking declared options, and Score for predicting levels on an ordered rubric.
Under the hood, separate state and action encoders each pair a frozen Qwen3-8B backbone with a trainable 20-million-parameter projection. A bidirectional InfoNCE objective draws a state toward the action taken and away from alternatives. At inference, dot products feed a softmax distribution, providing one primitive for best-of-N selection, tool routing, and other ranking tasks.
Training proceeds in three stages: roughly 60 million Nemotron DQA question-answer pairs, 30 million synthetic hard negatives generated with Gemini 2.5 Flash-Lite, and one million agent trajectories from Agent Data Protocol, Endless-Terminals, and LiteCoder-Terminal-SFT. On approximately 100,000 held-out questions, pretraining alone reaches 52.1% top-1 accuracy, and hard-negative mid-training raises that to 69.2%. Introducing hard negatives from the start peaks at 62.4% before overfitting, according to the project’s evaluation.
Latency is the release’s main attraction. With three actions, cached vectors reportedly reduce revisited-state processing on an RTX 4090 from 1.7 milliseconds to 0.6, while the model card claims a 13-fold advantage over Jev with approximately 1,000 candidates. The headline “up to 9×” figure comes from T-Rex, where actions recur across states, and CLM matches Jev on T-Rex and Super Mario. It trails Jev on tool calling and WikiRacing, although it reportedly runs faster on every tested task.
As a coding-agent verifier, CLM-8B also shows strong but bounded results. Fine-tuned heads record scores of 81.6% on 38 held-out DeepSWE tasks and 87.6% on 30 held-out Terminal-Bench 2.1 tasks, using best-of-four and best-of-five candidates attributed to Opus 5 and Fable 5, respectively. The team reports new state-of-the-art verifier results, with CLM running 4.1 to 5.7 times faster than Jev on an H100; Jev scores below pass@1 on both benchmarks. These are held-out subset results using fine-tuned heads, not full leaderboard submissions or zero-shot measurements.
For developers, the combination of an open model, a lightweight head, and a familiar API could make learned action scoring practical for private, self-hosted agents. Caching is especially relevant when state changes each turn but available tools remain stable, while weaker tool-calling results caution against equating speed with quality. Independent replication, broader hardware testing, and full-benchmark evaluations will determine how much of the claimed advantage survives outside project-controlled tests. Longer term, CLM-8B could support specialized verifier heads, but its credibility will depend on demonstrating robust decision quality alongside faster inference.
