A 364M Hebrew decision encoder on NeoDictaBERT: give it a state, a question and 2 to 7 free-text options and it returns a probability for each option in one forward pass, about 70 ms on a laptop CPU. Hebrew only.
Nandakishor M's typed-decision model. A frozen Qwen3.5 backbone is read, never trained, and cut off mid-stack; a small trained physics engine settles each choice, score or yes/no question into calibrated probabilities, a conformal set and an abstain flag. 0.8B and 4B.
Decides
choice, classify, route
choice, score, noul, classify, route
Architecture
nativ
vega
Fine-tuned from
dicta-il/neodictabert
qwen/qwen3.5-0.8b
License
cc-by-4.0
apache-2.0
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
96.2%
55.6%
Calibration error
0.010
0.151
Valid action rate
—
—
Median latency
—
—
p95 latency
—
—
Figures are from each model’s manifest; accuracy and latency are what the publishers report, on their own suites and hardware. Add a third model.
Questions
What is the difference between nativ-he-decision and vega?
nativ-he-decision is from Yoav Pinto and vega from Nandakishor M. Both have open weights you can download and run. Both answer choice, classify and route questions. Only vega answers score and noul. vega reads up to 72K tokens of state, against 1K tokens for nativ-he-decision. nativ-he-decision is the smaller model, at 364M parameters to 800M. nativ-he-decision is licensed cc-by-4.0; vega, apache-2.0.
Which is more accurate, nativ-he-decision or vega?
They report on different suites — nativ-he-decision 96.2% on nativ-bench user-intent task, 4 options (MASSIVE Hebrew, 2,973 items; the maker's own suite), vega 55.6% on His 11 October 2026 benchmark, published with its results and its harness (a notebook in vegaml): macro accuracy over 43 public tasks for the 0.8B, 0.717 for the 4B; with JevBench (231 items) and DecisionBench (1,075 rows), 3,841 decisions in all, seed 20261011. On his own set of 2,050 choice decisions over six intent workflows (esci, clinc150, appliances, banking77, when2call, apibank), one choice question each, which he has not published with its question wording, he reports 0.763 (calibration error 0.026) for the 0.8B with its public_intents adapter, 0.389 without the adapter and 0.803 for the 4B with it. He publishes no accuracy for the typed_decisions adapters — so the numbers do not rank them. Test both on your own labelled examples.
Which is cheaper, nativ-he-decision or vega?
nativ-he-decision: Free (open weights). vega: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run nativ-he-decision or vega locally?
Yes, both: systemone pull yoav-pinto/nativ-he-decision and systemone pull nandakishorm/vega download the weights.
Evaluation suite
nativ-bench user-intent task, 4 options (MASSIVE Hebrew, 2,973 items; the maker's own suite)
His 11 October 2026 benchmark, published with its results and its harness (a notebook in vegaml): macro accuracy over 43 public tasks for the 0.8B, 0.717 for the 4B; with JevBench (231 items) and DecisionBench (1,075 rows), 3,841 decisions in all, seed 20261011. On his own set of 2,050 choice decisions over six intent workflows (esci, clinc150, appliances, banking77, when2call, apibank), one choice question each, which he has not published with its question wording, he reports 0.763 (calibration error 0.026) for the 0.8B with its public_intents adapter, 0.389 without the adapter and 0.803 for the 4B with it. He publishes no accuracy for the typed_decisions adapters
