Contrastive-LM Releases CLM-8B: An Open System One Model That Scores Agent Actions Up to 9× Faster Than Jev
MarkTechPost Michal Sutter
CLM-8B is a new open model that scores actions instead of writing text. It’s meant to compete with Jev, and it’s faster in some agent loops.
Based on reporting by MarkTechPost, Michal Sutter — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Contrastive-LM has released CLM-8B, which it describes as the first open model in a new category: Contrastive Language Models. Unlike a normal chatbot, it doesn’t generate an answer directly. It scores candidate actions against the current state and turns those scores into probabilities. That puts it in the same lane as Jev, the proprietary System One model from TypeSafe AI.
The model is set up to be deployable. The Apache-2.0 head weighs 75 MB, runs on a single NVIDIA GPU under Linux, and uses vLLM to serve the Qwen3-8B encoder. The GitHub repo exposes a TypeSafe-compatible API with three request types: Noul for truth probabilities, Choice for picking from a declared set, and Score for ordered rubrics. Requests written for TypeSafe’s API can also be replayed through CLM’s Python client.
Under the hood, CLM trains a state encoder and an action encoder with a bidirectional InfoNCE loss. Both use frozen Qwen3-8B backbones plus 20M-parameter trainable projection heads. At inference time, the system scores each candidate with a dot product between state and action embeddings, then applies softmax. The result is useful for best-of-N ranking, tool routing, and typed decisions. The serving stack also caches vectors in a slab of GPU memory so repeated states get cheaper to process.
That caching shows up in the timings. On one RTX 4090 with three actions, revisited states drop from 1.7 ms to 0.6 ms. The model card says CLM runs 13× faster than Jev with about 1,000 candidates, while the paper’s zero-shot table shows up to 9× lower latency in the T-Rex game. In that table, CLM matches Jev on T-Rex and Super Mario, but trails on tool calling and WikiRacing while still running faster on every task.
Training was staged as well: about 60M Nemotron DQA question-answer pairs first, then about 30M synthetic hard negatives from Gemini 2.5 Flash-Lite, then about 1M agent trajectories from Agent Data Protocol, Endless-Terminals, and LiteCoder-Terminal-SFT. On around 100K held-out questions, pre-training alone reached 52.1% top-1 accuracy, mid-training raised that to 69.2%, and starting with hard negatives too early peaked at 62.4% before overfitting.
The verifier results are probably the sharpest part of the release. On held-out DeepSWE and Terminal-Bench 2.1 subsets, fine-tuned CLM heads beat Jev on pass@1 while running much faster on an H100. The team calls those new SOTA verifier results, though they are not full leaderboard submissions. For a model that refuses to talk and only wants to score, that’s a pretty strong argument.
My take — AI-written commentary, not fact-checked reporting
This is the kind of unglamorous AI work that actually matters: less chatbot cosplay, more decisions per second. Open models that can slot into agent loops without drama are a better bet than glossy closed systems that only look magical in a demo. The industry keeps worshipping generation, while the useful stuff is often just fast scoring and a good API.
Read more about this at: MarkTechPost