Organization repository on GitHub
- Forks
- 256
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Models & Reimplementations
CLM
Contrastive-language-model decision service exposing CLM-8B over a TypeSafe-compatible API: typed Choice, Score and Noul answers plus candidate ranking, with state and action embeddings cached separately so repeat candidates cost one dot product.
- Stars
- 3k
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Author
- Contrastive-LM
- Last verified
- Oct 2026
About this project
A System One Model for Fast and Generalizable Decision-Making. 🔥 Contrastive Language Models (CLMs) are a new class of System One model trained with a contrastive learning objective that connects states and actions. This repo serves CLM-8B behind a TypeSafe-compatible API. CLM-8B is pre-trained on 60M Nemotron Q&A pairs, mid-trained on 30M synthetic hard negatives, and post-trained on 1M agentic trajectories.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Two encoders, each a frozen backbone plus a 20M-parameter projection head, cache state and action embeddings independently.
- Typed questions return real distributions: Choice probabilities, ordered Score levels, and a Noul probability per statement.
- An in-process Engine ranks free-form candidates and answers typed questions against a pooling embeddings endpoint, with no server.
- The comparison suite runs on 38 held-out DeepSWE and 30 held-out Terminal-Bench tasks, with verifier latency measured on an H100.
Watch out
Running it needs a vLLM pooling server for Qwen3-8B embeddings plus a CUDA GPU, and the first clm-serve run pulls a 75 MB reference head. States over 2048 tokens are truncated unless both length limits are raised together, which costs more GPU memory. The README's data-recipe and scaling-law sections are cut off in the captured text, so training hyperparameters are not documented here. Accuracy and speed figures are the authors' own, including the comparison runs against Jev. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
It is a separate model, CLM-8B, served behind an API shaped like TypeSafe's System One endpoint. Questions may be passed as plain wire-format dicts, so a request written for Jev replays as client.system_one(state, questions). The latency and accuracy comparisons are the authors' own measurements, not an independent check.
Two processes. A vLLM server exposes Qwen3-8B as a pooling embeddings model on port 8090, and clm-serve runs on 8700 and downloads a 75 MB reference head on first run. A web playground is served at / and can be turned off with --no-ui.
States longer than 2048 tokens are truncated. To keep more, raise --max-model-len on the vLLM side and --max-tokens on clm-serve together, which the README notes needs more GPU memory. Option texts are embedded once and cached, so a cold cache bills more input tokens than a warm one.
Yes. docs/FINETUNING.md and train/finetune.py cover it, with a choice task pointed at a typed-decisions dataset and a holdout flag for task-disjoint splits. The captured README truncates the data-recipe and scaling-law sections, so the exact hyperparameters for reproducing a published number are not shown.
A state encoder and an action encoder, each a frozen LLM backbone with a 20M-parameter trainable projection head, trained with a bidirectional InfoNCE contrastive loss. CLM-8B is pre-trained on 60M Nemotron Q&A pairs, mid-trained on 30M synthetic hard negatives, and post-trained on 1M agentic trajectories.
As a verifier on agentic coding. On held-out sets it reports 81.6% on DeepSWE (31 of 38 tasks) and 87.6% on Terminal-Bench 2.1 (30 tasks), while Jev scored below pass@1 on both and ran 4.1 to 5.7 times slower. Those latency figures come from an H100, and the up-to-9x zero-shot claim is the authors' own measurement.
You host the whole thing, so there are no per-call API charges, just GPU. A typical call reports 38 input tokens and 58.1 ms once the option-text cache is warm, and 106 tokens on a cold cache, because option texts are embedded once and reused. Remote use is an SSH tunnel or clm-serve --no-ui for API only.
More Jev resources like this
1.7kGitHub stars
Open-weight 4B and 9B models from TokenRhythm: NeoHorse-1 for tool use, coding, and instruction following, plus NeoHorse-Jev-4B for prefill-only Choice, Noul, and Score decisions.
Jev Models & Reimplementations#community#jevbest
Community
TokenRhythm/NeoHorsePython
TokenRhythm/NeoHorsePython1.5kGitHub stars
A 0.8B decision model that answers first and escalates only when unsure, plus nine LoRA adapters of roughly 41 MB each for guard, triage, tool choice, spam, and similar classification work.
Jev Models & Reimplementations#community#jevbest
Community
firelex/jeffPython
firelex/jeffPython2kGitHub stars
DeepOpen is a non-autoregressive decision engine built on Laya that classifies into developer-defined types in one forward pass, with a language router picking between English, multilingual, and typed-decision checkpoints.
Jev Models & Reimplementations#community#jevbest
Community
deepopen-com/deepopenPython
deepopen-com/deepopenPython715GitHub stars
Decision models that answer several choice, score, and noul questions over a state with probabilities in one forward pass, with training, two inference backends, calibration tooling, and the benchmark suite in the repo.
Jev Models & Reimplementations#community#jevbest
Community
InternLM/Intern-DecisionPython
InternLM/Intern-DecisionPython402GitHub stars
Decision models that return typed probabilities for candidate labels from a state, instructions, and options, shipped as LoRA adapters plus a trained decision head over pinned Qwen base weights.
Jev Models & Reimplementations#community#jevbest
Community
Zefan-Cai/Open-JevPython
Zefan-Cai/Open-JevPython363GitHub stars
Open-weight 2B, 4B and 9B decision models that extend Jev's typed question format to photos, so an image plus a record can be checked against options you supply, with a probability on each and an explicit can't-tell. Runs on your own hardware.
Jev Models & Reimplementations#community#jevbest
Community
mohit67890/imajevPython
mohit67890/imajevPythonFrom the Jev guides
Original write-ups on what this entry is for.
Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.
Jev Models & Reimplementations#multimodal#open-models#decision-model
Read article
One model in September, a dozen by October. Clef, Liquid d1, Perplexity, and OpenAI all ship the same typed-answer shape now. How to read the rival benchmarks and pick on your own task.
Jev Models & Reimplementations#decision-model#benchmarks#pricing
Read article
Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.
Jev Models & Reimplementations#open-models#hugging-face#local
Read article
From the Jev community
Posts from builders shipping with Jev right now.