Developers building AI agents often hit a wall when they need a fast, typed decision component that can score candidate actions against a changing state without generating text. Proprietary System‑One models like Jev deliver the required interface but come with high latency, licensing restrictions, and no easy way to run them on modest hardware. This slows down agent loops, inflates cloud costs, and blocks teams that want to experiment, fine‑tune, or embed the verifier directly into their pipelines.
Contrastive‑LM’s CLM‑8B removes these barriers. It is an open‑weight model released under Apache‑2.0 that scores a set of candidate actions and returns probabilities, exactly matching Jev’s API. The head is only 75 MB, the encoder is a frozen Qwen3‑8B backbone, and the whole system runs on a single NVIDIA GPU under Linux using vLLM for serving. Because state and action encoders are separate, the action vectors can be cached and reused across steps, cutting latency from ~1.7 ms to ~0.6 ms on an RTX 4090 when states repeat—over a 2× speedup. On benchmarks CLM‑8B is 9× faster than Jev on the T‑Rex game and 4–5× faster on coding‑agent verification tasks while staying within a few percentage points of Jev’s success rate.
To put CLM‑8B to work:
- Pull the model from Hugging Face (CLM‑v0.1‑8B) and install the provided Python client.
- Start the server with
clm-serve --model CLM-v0.1-8B --backend vllm. - Send a JSON request that mirrors the TypeSafe format: include the current state description and a list of candidate actions.
- Receive a probability distribution over actions; pick the highest‑scoring option or use the scores for ranking, tool routing, or verification.
- If higher accuracy is needed for a specific domain, fine‑tune only the 20 M‑parameter projection heads using the three‑stage recipe (pretrain on QA pairs, add hard negatives, then finish on agent trajectories). The encoder stays frozen, so training is fast and cheap.
The result is a low‑latency, open, deployable verifier that can replace Jev in agent loops, cut inference costs, and let teams iterate on their AI systems without being locked into proprietary software.