Contrastive-LM Releases an Open Model for Faster Agent Decisions, With an Accuracy Tradeoff

An AI agent does not need to write an answer every time it chooses what to do next. Contrastive-LM has released CLM-8B to score supplied actions and return probabilities instead of generating text. In reported tests against TypeSafe AI’s Jev, it was faster across four tasks but less successful on two. The release offers a sharper choice than a headline speedup suggests: how much decision quality can an agent afford to trade for time?

A decision, not another response

CLM-8B takes the agent’s current state and a set of possible actions, then scores the match between each pair. It represents states and actions as vectors—numerical descriptions the model can compare—rather than writing a fresh explanation of each option. The resulting scores become probabilities for the available choices. This makes it a decision component for an agent, not a replacement for a model that needs to draft text or code.

Under the hood, separate encoders for the state and the action use a frozen Qwen3-8B model and trainable projection heads. Training draws matched state-action pairs closer together and pushes mismatched pairs apart. When an agent keeps choosing among the same actions, their vectors can be reused instead of computed again at every step. That reuse helps explain why the design is aimed at repeated choices such as routing a tool call.

Where the speed claim holds—and where success slips

The largest reported gap came in a T-Rex game test: CLM-8B took 16.5 milliseconds per decision versus 149.8 milliseconds for Jev, with both succeeding in all five trials. That is the source of the roughly ninefold speed figure. The models also tied on success in a Super Mario test, although CLM was faster there too.

Other tasks make the tradeoff harder to dismiss. On WikiRacing, CLM completed 26 of 30 tasks at 79.8 milliseconds, while Jev completed all 30 at 225 milliseconds. The tool-calling results show a smaller but similar split:

A different test for choosing code

The team also tested CLM as a verifier: another model generated several possible solutions, and CLM selected one. With task-specific fine-tuning, it achieved 81.6% on a 38-task DeepSWE subset and 87.6% on a 30-task Terminal-Bench 2.1 subset. Jev scored 71.1% and 83.1% on those respective subsets. CLM’s reported selection times were lower as well: 79 versus 449 milliseconds on DeepSWE, and 32 versus 131 milliseconds on Terminal-Bench.

Those coding figures do not describe the released zero-shot checkpoint: they use separately fine-tuned heads, and the evaluations cover held-out subsets rather than full benchmark submissions. They show that CLM can be trained to select among generated answers on these tests, not that its speed and accuracy carry over unchanged to every agent workflow.

What developers can run

The release includes a repository and a serving setup with a TypeSafe-compatible API. Its Python client supports three kinds of bounded question: judging whether a statement is true, choosing from listed options, and scoring against an ordered scale. That gives developers a way to try CLM in decision points shaped like Jev requests, rather than rebuilding the surrounding workflow first.

The released projection head is described as 75 MB and Apache-2.0 licensed; the serving setup uses the Qwen3-8B encoder and runs on one NVIDIA GPU under Linux. That is a concrete self-hosting route, though not a claim that the entire system occupies just 75 MB. The question for a deployment is whether its own mix of repeated actions and costly mistakes looks more like CLM’s tied game results or its lower-success tool and navigation results.