Skip to content

JevAny v0.3.0 · Model Family Release

Latest
Latest

Choose a tag to compare

View all tags
@weitianxin weitianxin released this 30 Sep 05:20
v0.3.0
75aa500

Model family

Five LoRA SFT checkpoints are now available under the SimpleJev organization:

JevAny-Qwen3.8-27B is the default and highest-accuracy release: 85.76% Transfer and 90.48% JevBench. The Qwen3.5 4B direct-token release reaches 80.95% JevBench; the pointer 4B release reaches 80.09% and supports larger choice sets.

Evaluation and data

  • The README now compares JevAny with Kev-4B, Kev-27B, Jev 1.13.0, and Laya using the same Transfer and JevBench presentation.
  • Machine-readable results include exact unrounded metrics, suite hashes, per-suite breakdowns, and released-checkpoint compute accounting.
  • Current training scale is 1,772,725 text records and 2,180,242 labelled decisions. Only aggregate scale and broad task families are documented.
  • The release family supports language and compatible native multimodal backbones.

Documentation and roadmap

  • Added the JevAny technical report covering architecture, readout tradeoffs, ablations, compute, limitations, and open research directions.
  • Updated training, inference, deployment, data, evaluation, demos, cases, and robotics documentation for the current model family.
  • Historical demo recordings retain their original checkpoint identifiers for provenance.
  • Full-parameter SFT, improved post-training, and hyperparameter-optimized variants remain planned releases.

Full Changelog: v0.2.0...v0.3.0

Assets 3
Loading