[AINews] Here are 6 Clones of Jev in 2 days

Six clones and reproductions of the Jev decision model appeared within two days of its launch, according to an AI news roundup. Jev launched on Wednesday and its launch video drew 36M views in two days, compared with 74M views for OpenAI's Navier Stokes result and 57M for Anthropic's Fable 5. The roundup states Jev was not open source, which it says invited speculation, demos, examples and criticism.

The reported reproductions include Laya, described as a 421M-parameter ModernBERT-large encoder with two added transformer layers that score user-supplied options, using PPO over sequence embeddings to output turn-by-turn conversion trajectories as probabilities from 0.0 to 1.0. DiffusionGemmaJev takes a diffusion model basis and is described as pretty close on benchmarks. Bespoke Nimble is a LoRA fine-tune of Qwen3.5-9B using synthetic contrastive data curation, reported close but slightly lower on benchmarks. SemIf (formerly OpenJev) uses 4B and 35B causal Qwen3.5 backbones with a tiny three-class NLI classifier on the last token. Jevlike is a 40K byte embedding lightweight option-attention model in which each candidate becomes a query reading from a shared context representation. Kev-0.5B is a LoRA adapter plus a small readout head on top of Qwen2.5-0.5B.

On the data side, the roundup says the training data is acknowledged to be 100% synthetic.

The roundup also reports that Jev is available as an eval model in Braintrust, with roughly 400x lower scoring cost versus prior setups, and that Bespoke Nimble's base Qwen improved from 66% to 90% on its curated eval versus 93% for Jev, with a reported 100ms on H100. The roundup states there is still no standard benchmark for this category.