CommunityJev Models & Reimplementations
Jev-Style 2B Decision v3 GGUF
The v3 Jev-Style decision model at 2B in GGUF: F16 3.78 GB and Q8_0 2.01 GB match PyTorch FP32 on 100% and 99.7% of parity rows, 73.6% on JevBench v1.4.1 public items.
- Published by
- Community
- Author
- chaoliangUNSW
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Three files built with unmodified llama.cpp at commit 441df11f: F16 3.78 GB, Q8_0 2.01 GB and Q4_K_M 1.27 GB.
- F16 is 100% top-1 identical to PyTorch FP32 and Q8_0 99.7%; Q4_K_M holds 95.7% with a 0.346 maximum probability delta.
- 73.6% on JevBench v1.4.1 public items, the highest among Qwen3.5-2B-family systems on the board, with 95% CI 67.6-78.9%.
- 25,600 tokens per call with no option cap; questions and options over 2,048 tokens go through a numbered-option catalogue, nothing truncated.
- Reads a 24,501-token state once: the first question takes 16.2 s and a further question about the same state 0.17 s.
Watch out
The bundled jev-score-v2 scorer must be built once from build_jev_score.sh, and the jev-style package 0.3.0 or later is required. The 73.6% is self-run with the official harness rather than a board entry, and for long documents the card says to use Q8_0 or F16 because Q4_K_M is noticeably noisier.
Questions about this entry
Short answers drawn from the linked project or page.
Install the jev-style package 0.3.0 or later, build the jev-score-v2 scorer once from build_jev_score.sh, then run jev-style serve --release 2b --backend gguf --quant Q4_K_M --scorer /absolute/path/to/jev-score-v2, which serves a local /v1/systemone API.
Q8_0 or F16. The card warns that Q4_K_M is noticeably noisier on long inputs, even though it passed its accuracy gate and handled 42 of 43 long-fixture questions.
73.6% on the 231 public items, self-run with the official harness rather than an official board entry, with a 95% Wilson interval of 67.6-78.9%. Jev 1.13 scores 86.6% on the same board and decider-2b's 71.0% lies inside the interval.
25,600 tokens shared by state, questions and options, with no option cap. Questions and options above 2,048 tokens use a numbered-option catalogue; the long fixture reached 151 options, and the card says nothing is truncated.
On an M1 Max with the F16 file, the first question about a 24,501-token input took 16.2 s and a further question about the same state 0.17 s; ten questions in one call took 16.8 s. The state is computed once and reused.
Converted and quantized with unmodified llama.cpp tools at commit 441df11f, with gates declared before any format was scored: top-1 agreement for F16 and Q8_0, and a maximum accuracy drop of 1.0 point for the 4-bit file.
On the main Jev-Style-2B-Decision-v3 model card, which this repository points to for full results, protocols, training data and licences. This GGUF repository ships LICENSE and NOTICE files alongside the three quantized builds.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: ASTGL replay chose local Laya routing over hosted Jev on one Mac Studio setup.
Jmeg8r15 pointsDoes an open-weight decision model beat a hosted one? Jev vs. LayaRead the thread on Hacker NewsReddit: A walkthrough of routing between models with Jev.
Facebook: Intake Router keeps diagnosis out of the answer space at the front desk.
FacebookClinicnotes
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
More Jev resources like this
Mithril's open, Jev-shaped typed-decision model on DeBERTa-v3-large: choice up to 255 options, score and noul from one forward pass, 0.854 in-domain with ECE 0.022, trained on public gold labels only.
Jev Models & Reimplementations#community#open-models#local
Community
The GGUF edition of TokenRhythm's NeoHorse-Jev-4B: BF16 through Q2_K quantizations that each bundle the language backbone, visual weights, decision head, and tokenizer, plus a shared runtime for text and single-image requests.
Jev Models & Reimplementations#community#open-models#local
Community
ggmlc-compiled GGUFs of the Laya typed-decisions specialist on ModernBERT-large: F16, Q8_0, and UD_Q4_K_M files for invoice, security, customer-service, and harness presets, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
A 70.8M-parameter DeBERTa-v3-xsmall cross-encoder that answers typed decision questions in one batched pass with zero generated tokens, shipped with a calibration.json that the engine requires.
Jev Models & Reimplementations#community#open-models#local
Community
A decision model that scores its own clarifying questions by value of information: at AUC .799 against an oracle .797, it asks only when asking is expected to change the answer, then decides or hands off.
Jev Models & Reimplementations#community#open-models#local
Community
GGUF builds of the Jev-Style v3 decision model on Qwen3.5-0.8B: a 0.53 GB Q4_K_M file the card says is decision-identical to FP32 on 240/240 parity rows, served through a pinned llama.cpp scorer.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.
Jev Models & Reimplementations#open-models#training#local
Read article
Interface, size, license, claims — in that order. A selection guide across the system-one catalog, from a 706K form scorer to 27B Jev students.
Jev Practices & Patterns#open-models#evaluation#benchmarks
Read article
From the Jev community
Posts from builders shipping with Jev right now.