Jev directory
CommunityJev Models & Reimplementations
APUS-OpenJev (APUS Reflex)
APUS AI Lab's decision model family for browser agents and business workflows, in 4B, 9B and 35B-A3B sizes with a low/high effort switch. On the lab's frozen 80-question panel the 35B-A3B scores 88.75% against 82.5% for the Jev API.
- Reading time
- 2 min
- Published by
- Community
- Author
- APUS AI Lab
- Last verified
- Oct 2026
- Platforms
- API
- Capabilities
- Open sourceRuns locally
Highlights
- 4B, 9B and 35B-A3B variants share one decision interface; applications pick effort low or high to trade compute for quality.
- Frozen 80-question panel: 35B-A3B 71/80 (88.75%), 9B 68/80 (85.0%), 4B 66/80 (82.5%), Jev API 66/80 (82.5%), Laya 55/80.
- The optimized 9B service reports 25.58 ms median HTTP latency in the supplied snapshot, against 280.80 ms pooled for the Jev API.
- The low path pairs an earlier trained exit with candidate-aware output projection that scores only the supplied candidate labels.
- Branding note: the family is now APUS Reflex; some filenames and served-model aliases keep historical identifiers for compatibility.
Watch out
Model card reviewed 2026-10-15; the fetched text was truncated inside the latency section. The 80-question panel is the lab's own development set and the card says the differences do not establish broad or statistically significant superiority; the 35B row uses checkpoint-5949 while the published 9B accuracy uses checkpoint-3000.
Questions about this entry
Short answers drawn from the linked project or page.
A family of decision models from APUS AI Lab for browser agents and business workflows. Given a task, context and candidate actions, it returns a scored choice an application can execute, in 4B, 9B and 35B-A3B sizes with a common interface.
On the lab's frozen 80-question panel the 35B-A3B scores 88.75% and the 9B 85.0%, versus 82.5% for the Jev API and 82.5% for the 4B. The card states these panel differences do not establish broad or statistically significant superiority.
Applications choose effort=low or effort=high. The native low path pairs an earlier trained exit with candidate-aware output projection that computes scores only for the supplied candidate labels, reducing executed depth and output-head work.
With joint post-training across computation depths: both execution paths learn from reference decisions, and the complete path supplies a distribution-level learning signal for the shorter one, so early decisions are trained rather than untrained.
The supplied snapshot reports 25.58 ms P50, 222.36 ms P95 and 275.58 ms P99 on one RTX PRO 6000 Blackwell 96GB GPU with 333/333 successful requests and 0% errors. That is full-response latency, not first-token latency.
A branding note says the family is now APUS Reflex and that some weight filenames, script and module names, served-model aliases and the evaluation dataset ID keep historical identifiers for download and runtime compatibility. They do not indicate different model versions.
It is the lab's own frozen development panel, with 16 questions per task family where one answer moves accuracy by 6.25 percentage points. Repetition across runs establishes consistency on these questions, not additional independent evidence.
More Jev resources like this
cklxx fine-tuned Laya into the decision head of browser-use/jev-ultrafast: one bidirectional encoder pass per step picks the browser operation and target element, and the server answers Jev's /v1/systemone requests identically.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
A 144.3M-parameter bilingual typed decision model fine-tuned from jhu-clsp/mmBERT-small: it answers noul, choice, and score questions in one forward pass with calibrated confidence and zero output tokens, and it runs on CPU.
Jev Models & Reimplementations#community#catalog-import#cobanov
A 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 Japanese System One decision model built on modernbert-ja-310m: send text and typed questions and get typed answers with probabilities from one forward pass, no text generation, at 314.6M parameters that also run on a CPU.
Jev Models & Reimplementations#community#catalog-import#cobanov
Community
hiroki-abe-58/sokudanPython
hiroki-abe-58/sokudanPythonAn open-source, Jev-compatible System One model: Dex answers choice, score and yes/no questions over text in a single forward pass from an Ettin-1B encoder, and speaks the same /v1/systemone shape as Jev.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
canadaornaw/dexPython
canadaornaw/dexPythonGestalt Labs' Jeff is a LoRA adapter on Qwen3-4B-Instruct-2507 that answers Jev-style choice, noul and score questions with labels and probabilities instead of prose, runs locally, and ships a train-your-own guide; on 9,730 fact-checking examples it trails Jev 1.13.0 on accuracy (0.8183 vs 0.8283).
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
From 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.