QEV-1.3B (QEV-2B-compact)

QEV-1.3B was published as QEV-2B-compact; the old repository id redirects here.

A compact multimodal decision model with 1,314,213,824 parameters (59% of QEV-2B). Give it evidence (text and up to four images) and candidate options; it returns a probability for each option, a typed answer and a separate abstain flag in one forward pass, without generating tokens.

QEV-1.3B is QEV-2B v3 with the language model cut after block 16 of 24 and the input vocabulary trimmed to 32,768 rows. A decision readout needs no output vocabulary, and our depth probe found the decision complete after block 16, so the removed parts add nothing to the decisions.

Results

Public half of the JEV Decision Index vision suite (12,210 rows, 11 benchmarks, public kit, official qev_decision_index:QEVEngine loading this repository from the Hub at a pinned commit). These are our own runs; the board's private half is run by the maintainers, so the board's full score for this model is not known yet.

Model	Parameters	Public vision index	95% bootstrap interval
QEV-1.3B	1.31B	62.15	60.7 to 63.5
QEV-2B v3 (same kit and engine, our run)	2.21B	62.27

The compact model is −0.12 points from its parent on the same rows (paired item bootstrap, 95% interval −0.90 to +0.70). Run records are in ken-jo/decision-index-results. Per benchmark (public skill points; weight in brackets):

Benchmark	QEV-1.3B	QEV-2B
CV-Bench (1)	71.1	71.8
BLINK (1)	48.1	48.0
RealWorldQA (1)	56.2	55.3
CharXiv (1)	65.2	65.7
InfographicVQA (1)	80.9	81.4
Mind2Web (1)	81.8	83.6
Winoground (1)	66.0	65.5
KIE, CORD+FUNSD (1)	92.3	92.6
Moderation, Hateful Memes (1)	31.4	31.6
R-Bench-M (0.5)	19.5	15.4
MMMU-Pro vision (0.25)	13.2	16.0

Speed and memory

One RTX 4060 Ti 8 GB, batch 1, bf16, 47 request profiles of up to 16 options at full resolution (1.6 MP per image), 3 warm-up and 30 timed runs per profile, with a control run of the previous QEV-0.8B release in each session.

Model	Worst-case p95	Median p50	Peak GPU memory
QEV-1.3B	630.4 ms	99.0 ms	2.76 GiB
QEV-2B v3, accurate mode	736.4 ms	122.9 ms	4.44 GiB

The compact timings come from a second session the same day (control within 10% of its published values). They were measured on these exact weights through our evaluation loader; the release loader gives identical outputs on all 900 development questions.

Calibration and abstention

Temperatures and abstention thresholds were refitted for this package on the 3,000-row release Calibration DEV, with the same procedure as v3: on group-disjoint halves, the temperature minimises NLL on one half, and the abstain threshold on the other half takes the highest coverage with at most 15% empirical error and at least 20 accepted examples. No public-suite row and no holdout row was used.

Type	Temperature	Abstain below	Accuracy (policy half)	ECE	Coverage
choice	1.029	0.301	89.5%	0.019	100%
score	0.579	0.454	88.6%	0.077	100%
noul	1.223	0.674	71.3%	0.070	61.3%

The 15% error target is an observed rate on our development data, not a guarantee, and it does not carry over to other data without refitting. An answer is still returned when the model abstains.

Architecture

• Vision encoder and projector: unchanged from QEV-2B v3 (frozen Qwen3.5 encoder).
• Language model: blocks 1–16 of 24 and the final normalisation, from the merged QEV-2B v3 weights; blocks 17–24 are removed.
• Vocabulary: 32,768 of 248,320 rows kept (every token produced on all training and development prompts, all special and byte tokens, and the merges needed to build them), with a rewritten byte-fallback tokenizer and vocab-map.json. Text outside these domains may take more tokens; over-budget requests are refused, never truncated.
• Readout: the block-16 option readout (four reasoning slots, one decision view, rank-64 conditioned binding head), distilled from the full-depth head on training questions and selected on development data only.
• Serving: accurate mode only (up to 1,638,400 pixels per image, 8,192 input tokens, at most 4 images). fast mode is not offered.

The checkpoint uses the bundled-backbone format (manifest.json format_version 3): every file under backbone/ is listed with its SHA-256, and the loader refuses missing, extra or modified files. Loading needs no download besides this repository.

Quickstart

pip install "qev @ git+https://github.com/ken-jo/qev@e091f7ae8c6f064860d59be560faa27f0387ff35"

from qev_decision_index import QEVEngine
engine = QEVEngine(model="ken-jo/QEV-1.3B")  # accurate mode, CUDA

or load the model directly with veyra.option_model.OptionModel.load(path, device="cuda") on a downloaded snapshot. See the qev repository for the request format.

Intended use and limitations

• Intended for typed decisions over given options where memory is limited: choosing among candidates, extracting fields from receipts and forms, picking screen elements for an agent, yes/no checks.
• Not for automated content moderation or any decision about people without human review (Hateful Memes skill is about 30), nor for medical, legal or financial decisions, free-form generation or languages other than English.
• Inherits the training data, contamination audit and failure modes of QEV-2B v3 (see its model card). Its private-half score is not known yet; its parent loses 14 to 16 points from the public to the private half of the board.
• Single training run; intervals cover item sampling only.

Citation

Technical report: QEV: calibrated single-pass decision models (Ken Jo, 2026), see ken-jo/decision-index-results.

License

Apache-2.0, as for QEV-2B and Qwen3.5. Some training sources have their own terms (for example FUNSD, non-commercial research; Hateful Memes, research use); the licence covers these weights and code, not those datasets.