Carve-Jev
Open Apache-2.0 Jev-like model family on Qwen3.8-27B that reads one-token candidate preferences and serves a TypeSafe-compatible API.
Overview
CARVE (Candidate-Aligned Readout via Vocabulary Evidence) is an independent, open-source Jev-like model family built on Qwen3.8-27B. It assigns candidate letters to user-defined options, scores only those letters and returns typed Noul, Choice or Score answers through a TypeSafe-compatible System One API, making each decision with one output token and no free-form JSON. It runs under vLLM with one-command setup.
The README is careful about semantics: CARVE probabilities are raw restricted candidate-token preferences, not calibrated probabilities of correctness, and every response labels its probability level. Public protocols, frozen revisions, raw summaries, ablations and negative results sit beside the code. In its published comparison CARVE scores 82.22%, second behind Jev's published 87.19%; Jev speed is deliberately omitted from the comparison.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Yes/no question over the state | yes · no |
| Choice | Which user-defined option fits? | — |
| Score | Where on the ordered scale? | — |
How it works
- Assign letters to options
- Run one forward pass on Qwen3.8-27B in non-thinking mode
- Read restricted candidate-token preferences
- Return typed answers via the System One API
Strengths
- No JSON parsing; one token per decision
- Honest about uncalibrated probabilities
- Reproducible evidence including negative results
Caveats
- Probabilities are not calibrated
- Needs a large GPU for a 27B model
Getting started
Follow the Quickstart for the one-command vLLM setup.At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- vLLM server
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Open-source Jev-like model family with one-token Qwen3.8 decisions, TypeSafe API, one-command setup, and reproducible benchmarks.
hellogumbo/awesome-jevOpen Jev-like model family that reads one-token decisions from Qwen3.8 under vLLM and serves a TypeSafe-shaped API, with one-command setup.
MrJev/awesome-jev
Frequently asked questions
Are CARVE's probabilities calibrated?
No, they are raw candidate-token preferences, as the README warns.
How does it compare to Jev?
82.22% versus Jev's published 87.19% in its comparison.
Is it affiliated with TypeSafe?
No, it is independent.
Related projects
Toy local Jev-shaped decision API on a 0.5B model that answers yes/no, choice and score questions by scoring option tokens.
ServiceprototypePythonSmall open decision models (0.8B and 4B, from Qwen3.5) with a server that speaks Jev's /v1/systemone format in one forward pass.
ServicebetaPythonServes MiniCPM5-2B-8bit on Apple Silicon as a Jev-compatible /v1/systemone service and reproduces its JevBench score.
ModelprototypePython- reflex★ 166
Small open decision model on a frozen Qwen3.5-4B serving /v1/systemone in about 200 ms, benchmarked against Jev on JevBench.
ServiceresearchPython Open MiniCPM5-2B decision model serving /v1/systemone: typed choice, noul and score probabilities in one forward pass.
ModelresearchPythonDocker recipes and ARM64 images for serving ten native decision checkpoints, such as Cloudflare Clef, behind /v1/systemone on DGX Spark.
ServicebetaPython