User repository on GitHub
A tree-based decision model on Qwen3.8-27B: shared context, isolated branches, and direct probabilities.
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Models & Reimplementations
Canopy-Jev 27B
An open decision model built on frozen Qwen3.8-27B that processes a shared application state with isolated question branches and returns probabilities with zero generated answer tokens. It reports 87.45% on JevBench, exceeding Open-Jev-27B-v1.1 by 2.16 points, and needs 64 GiB of free GPU memory.
- Stars
- 4
- Language
- Python
- License
- MIT
- Reading time
- 2 min
- Published by
- Community
- Author
- Camellia86
- Last verified
- Oct 2026
Used inJev for Software & DevTools
Highlights
- Reports 87.45% on JevBench and 75.68% on the Hard subset, exceeding Open-Jev-27B-v1.1 by 2.16 and 3.60 points.
- 6.26M trainable LoRA parameters; adapter and numerical prior total 25.24 MB on top of the frozen base.
- Choice, Noul and Score interfaces share one request-local prefix with isolated branches of up to 16,384 packed tokens.
- Zero-shot API P50 latency of 272-741 ms on a shared H200 across nine measured workloads; model load 37.58 s.
- MIT code with Apache-2.0 base-derived inference artifacts; MMLU-Pro reported at 64.50%.
Quickstart
bash
git clone https://github.com/Camellia86/Canopy-Jev.git
cd Canopy-Jev
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[cuda]'
hf download Camellia86/Canopy-Jev-27B --include 'artifacts/*' --local-dir .
canopy-jev doctorWatch out
README reviewed 2026-10-17; the comparison uses Open-Jev's published JevBench results, training used sampled Dirichlet policy gradients, and the model needs Linux with NVIDIA CUDA and at least 64 GiB of free GPU memory, so the numbers are not reproducible on ordinary hardware.
Questions about this entry
Short answers drawn from the linked project or page.
Canopy-Jev is an open decision model built on a frozen Qwen3.8-27B base that processes a shared application state and multiple isolated question branches, returning probabilities with zero generated answer tokens. Its weights are on Hugging Face and the code is on GitHub.
It follows the Jev-style System One pattern and is benchmarked on JevBench: the README reports 87.45% overall, exceeding Open-Jev-27B-v1.1 by 2.16 points, with 75.68% on the Hard subset and an 87.79% family mean. It compares against Open-Jev's published results.
Frozen layers 0-59 encode a tree of the shared state; copies of layers 60-63 carry the LoRA adaptation. Causal tree masks and branch-local positions isolate attention, prefix convolution and recurrent states are copied per branch, and a numerical prior reads branch endpoints to return fixed-policy probabilities.
Linux with NVIDIA CUDA and at least 64 GiB of free GPU memory, tested with Python 3.10, PyTorch 2.6.0/CUDA 12.4, Transformers 5.8.0, FLA 0.5.2 and Triton 3.2.0. The base model downloads separately from a pinned official snapshot.
Zero-shot API latency on a shared H200 is P50 272-741 ms and P95 282-762 ms across nine measured workloads, depending on state tokens and question count. Model loading measured 37.58 seconds separately, and timing includes tokenization, model, decoding and JSON encoding.
Original code and documentation are MIT; base-derived inference artifacts retain Apache-2.0. The repo ships LICENSE, WEIGHTS_LICENSE.md, NOTICE and THIRD_PARTY_NOTICES.md with complete terms and attribution.
Ruoxi Qiu publishes it under the GitHub user Camellia86; the repository was created 2026-10-02 and pushed the same day, with English and Simplified Chinese README sections and a CITATION.cff for the software entry.
More Jev resources like this
Three open-weights 4B decision models for English (Deem-4B), Turkish (Karar-4B) and German (Wähler-4B) that choose between options and return a probability for each. The server implements the Jev API, POST /v1/systemone, so existing Jev clients work by changing the endpoint.
Jev Models & Reimplementations#community#github#python
Community
mertkayacs/jevaltPython
mertkayacs/jevaltPythonA Chinese-language decision model that returns a probability for each fixed candidate option in one forward pass, with no generated text or chain of thought. It is a LoRA fine-tune of Qwen3.5-0.8B on 13,048,000 CJ-Dataset rows, benchmarked on CJ-Bench against Chinese-Jev.
Jev Models & Reimplementations#community#github#python
Community
yanqiangmiffy/chinese-decisionPython
yanqiangmiffy/chinese-decisionPythonA local yes/no decision model that speaks Jev's wire format on a laptop CPU, refusing Choice and Score until those question types ship.
Jev Models & Reimplementations#community#python#open-models
Community
feder-cr/jevC++
feder-cr/jevC++A System One layout optimizer for decision-model workloads: it reorders rows and fields so a prefix-caching backend reuses input computation, then returns per-record decisions in original order. Its bundled backend serves AutoTrust JEV-9B through vLLM.
Jev Repos & SDKs#community#github#python
Community
0814wdwd/solo_jevPython
0814wdwd/solo_jevPythonA small, fast decision model — a system one model — for agentic workflows from Strands Labs: it picks between options or rates on a scale faster than an LLM and returns a calibrated confidence with every decision.
Jev Models & Reimplementations#community#github#python
Community
strands-labs/strands-deciderPython
strands-labs/strands-deciderPythonSmall Jev-style decision models that read a state and typed questions and return a probability for every option in one forward pass without generating text; the 0.75B and 4.2B versions speak TypeSafe's Jev API format at a median of about 50 ms per request on one H100.
Jev Models & Reimplementations#community#github#decision-model
Community
jiwidi/jiwoPython
jiwidi/jiwoPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.