JEV_Humanoid
Repository serving CLM-8B, a contrastive System One model behind a TypeSafe-compatible API, aimed at humanoid robot decisions.
Overview
JEV_Humanoid is described in its list summary as using a contrastive learning model (CLM) for decision-making in humanoid robots. The README itself documents Contrastive Language Models, a class of System One model trained with a contrastive objective linking states and actions, and serves CLM-8B behind a TypeSafe-compatible API. Per the README, CLM-8B was pre-trained on 60M Q&A pairs, mid-trained on 30M synthetic hard negatives and post-trained on 1M agentic trajectories.
The README claims performance on par with Jev across computer use, gaming and tool calling with up to 9 times lower latency, and state-of-the-art verifier results after fine-tuning on Terminal-Bench 2.1 and DeepSWE; these are the authors' claims. States and actions are embedded separately and cached. Serving uses vLLM for Qwen3-8B embeddings plus a small reference head, and a Python client asks Noul, Choice and Score questions. The excerpt does not show humanoid-specific code.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this urgent? | — |
| Choice | Which team should handle this? | — |
| Choice | Which action should the robot take in this state?(inferred) | — |
How it works
- Serve Qwen3-8B embeddings with vLLM
- Start the CLM API, which downloads a small reference head
- Send states and typed questions through the client
- Use answers for agent or robot decisions
Strengths
- Open, self-hosted Jev-compatible API
- Cached state and action embeddings
- Same question types as Jev
Caveats
- Benchmark claims are the authors'
- Needs a GPU for the 8B encoder
- Humanoid application not shown in the excerpt
Getting started
`pip install contrastive-lm`, run `vllm serve Qwen/Qwen3-8B ... --port 8090` and `clm-serve`, then use CLMClient.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- vLLM plus CLM API server
- Has code
- yes
Use cases and integrations
How curated lists describe it
Using CLM (contrastive learning model) to decision making in humanoid robots.
wh000wh000/awesome-jev-live
Frequently asked questions
What model does it serve?
CLM-8B, a contrastive System One model with a TypeSafe-compatible API.
Is it Jev?
No, it is an open alternative that mimics Jev's interface.
What hardware is needed?
A GPU able to serve Qwen3-8B embeddings with vLLM.
Related projects
Mangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython- blink★ 22
Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.
ServicebetaC - Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaRuns the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.
ServicebetaTypeScriptllama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++