OpenJev (Verdict)

A different model. It is a 151M-parameter bidirectional ModernBERT-base encoder with a GLiClass bi-encoder classification head, 151,378,177 parameters in total, post-trained and described as inspired by TypeSafe's Jev and by RLCD rather than released by either. The accuracy table is scored on 231 public JevBench tasks, not on Jev itself.

No. The weights are byte-identical to the published heman10x/rlcd-modernbert-151m checkpoint. The three changes are inference-engine fixes: auto-loading calibrator.json and calibrating across all supported candidate counts instead of only 5, framing candidate labels as NLI hypothesis sentences so they match the backbone's pretraining, and cutting the context budget from 1024 to 512 tokens after training on states under 71 tokens.

Standard-tier accuracy rose from 62.5% to 69.4% and easy tier from 85.4% to 87.5%, but hard-tier accuracy across 111 tasks did not move at all: 36.9% before and after. What improved on that slice was calibration, with ECE falling from 0.298 to 0.118 and probability fidelity rising 10 points to 72.8.

The design targets a single non-autoregressive forward pass answering several typed questions in under 35 ms, with no token decoding and no JSON to parse. The README does not state the hardware, batch shape, or timing method behind that figure, so it cannot be checked or reproduced from the repository as documented.

Apache 2.0. The weights and artifacts are hosted on Hugging Face at heman10x/rlcd-modernbert-151m, and the full benchmark details and leaderboards are at Benchmark Heaven Jev Models.

It has 151,378,177 parameters: a bidirectional ModernBERT-base encoder with a GLiClass bi-encoder classification head. Document context and candidate labels are concatenated with [TEXT] and [LABEL] delimiter tokens and scored in one pass.

Candidate labels were previously evaluated as bare noun phrases. Because the GLiClass backbone descends from natural language inference formulations, formatting candidates as 'It is {description}' aligns inputs with pretrained representations and lifted easy and standard accuracy.