CLM
CLM-8B, a contrastively trained System One model that embeds states and actions separately, served behind a TypeSafe-compatible API.
Overview
Contrastive Language Models (CLMs) are a System One model class trained with a contrastive objective that connects states and actions. This repository serves CLM-8B behind a TypeSafe-compatible API and provides data, weights and a fine-tuning tutorial. CLM-8B was pre-trained on 60M Nemotron question-answer pairs, mid-trained on 30M synthetic hard negatives and post-trained on 1M agentic trajectories.
Because states and actions are disaggregated, their embeddings are cached and reused independently, which makes training and serving cheap and fast. The README reports performance on par with Jev across computer-use, gaming and tool-calling tasks with up to 9x lower latency, and, with light fine-tuning, verifier results of 87.6% on Terminal-Bench 2.1 and 81.6% on DeepSWE. It runs with a Qwen3-8B embedding server under vLLM plus the CLM API, and its Python client accepts Noul, Choice and Score so TypeSafe-written requests replay unchanged.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which action fits this state? | candidate actions |
| Yes / no | Is this urgent? | yes · no |
How it works
- Serve Qwen3-8B embeddings with vLLM
- Start the CLM API
- Send TypeSafe-shaped requests
- Cache state and action embeddings for reuse
Strengths
- Embeddings cached independently for speed
- TypeSafe-compatible requests
- Open data, weights and tutorial
Caveats
- Benchmark claims are the authors'
- Needs a GPU for the embedding server
Getting started
pip install contrastive-lm, then serve Qwen3-8B with vLLM and start the CLM API per the Quickstart.At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- vLLM + API server
- Has code
- yes
- Stars
- 3,014
Use cases and integrations
How curated lists describe it
Open alternative: an 8B System One model that answers the same `Choice` and `Noul` questions behind a TypeSafe-compatible API, matching Jev across computer-use, gaming and tool-calling with up to 9x lower latency and reporting 87.6% on Terminal-Bench 2.1 as a fine-tuned verifier.
yibie/awesome-jevContrastive Language Models (CLMs) are a new class of System One model trained with a contrastive learning objective that connects states and actions.
wh000wh000/awesome-jev-liveCLM-8B, a contrastively trained System One model served behind a TypeSafe-compatible API, with data, weights and a fine-tuning tutorial.
MrJev/awesome-jev
Frequently asked questions
How is it different from Jev?
It is open and contrastive, embedding states and actions separately so they can be cached.
Will my TypeSafe requests work?
Yes, the API and client accept the same question shapes.
How fast is it?
The README reports up to 9x lower latency than Jev.
Related projects
From-scratch research reconstruction of a small System-One decision model, used to route and gate commands in a Windows palette.
ProjectresearchPythonTokenRhythm's NeoHorse repo: open 4B/9B agent models plus NeoHorse-Jev-4B, a prefill-only Choice/Noul/Score decision model.
ServicebetaPythonOneJev: open multimodal System One model (0.8B-27B) answering typed questions about screenshots, photos, video and text in one pass.
ModelbetaPython706K-parameter Jev-style option scorer that rates fill, check, click or skip for each form element in one parallel pass.
ModelresearchPythonOpen infrastructure for training, evaluating and deploying System 1 decision models, with calibrated LoRA checkpoints on a 27B backbone.
ModelresearchPythonOpen Jev-style decision models with vision (Qwen3.5 4B/9B LoRA) serving Choice, Score and Noul on /v1/systemone.
LibraryresearchPython