# Jev clones are popping up everywhere just two days after launch > Source: > Published: 2026-09-19 16:46:32+00:00 # Jev clones are popping up everywhere just two days after launch The hype around Jev is actually wild. The launch video hit 36M views in two days, which is insane when you realize it's almost half of what OpenAI's Navier Stokes result got (74M) and way more than Anthropic's Fable 5 (57M). At my company, the team is already debating whether we can swap out some of our routing logic for these discriminative models. The big draw is that Jev isn't generative—it's a "System 1" decision model. One dev mentioned it's now an eval model in Braintrust with scoring costs roughly 400x lower than what we were using before. Since Jev isn't open source, the community has spent the last 48 hours trying to reverse-engineer it. We've already seen six different "clones" or approximations trying to hit those same benchmarks. Here is the current landscape of the Jev-likes: - **Laya:** This one uses a 421M param ModernBERT-large encoder. It adds two transformer layers to score options and uses PPO over sequence embeddings for turn-by-turn conversion trajectories (probabilities from 0.0 to 1.0). - **DiffusionGemmaJev:** Takes a totally different approach using a Diffusion model. Apparently, it's landing pretty close to the official benchmarks. - **Bespoke Nimble:** A LoRA finetune of Qwen3.5-9B. It uses contrastive data curation and constrained decoding. On curated evals, the base Qwen hit 66%, Bespoke Nimble hit 90%, and the real Jev is at 93%. It reportedly runs in 100ms on an H100. - **SemIf (formerly OpenJev):** Available on HF with 4B and 35B causal Qwen3.5 backbones. It uses a tiny three-class NLI classifier on the last token. - **Jevlike:** A lightweight option-attention model with 40K byte embeddings. It treats candidates as queries that read from a shared context and then assigns a score. - **Kev-0.5B:** The tiny option. It's a LoRA adapter and a small readout head on top of Qwen2.5-0.5B, meaning it can actually run on a MacBook Pro. [MCP](https://promptcube3.com/en/tags/mcp/)-style decisions away from small generative LMs and back toward discriminative models. It's basically a near-zero-marginal-cost judgment layer for things like UI adaptation or sensor-driven decisions. The only catch is the data. Everyone is talking about the architectures, but the data used for these is 100% synthetic. I'm curious to see if the "open" versions can maintain that accuracy without the proprietary synthetic sets Jev was trained on. For now, the fact that @typesafeai reached about 13% of teams—which is 2x the GPT-5.6 family and 6x Fable 5.1—shows that people are desperate for a fast, cheap way to handle routing and legal ops decisions without the overhead of a full LLM. [Next Why the Llama-3.1 animal-number test proves subliminal prompting isn't just about vector geometry →](https://promptcube3.com/en/threads/9507/) ## All Replies （2） I'm curious if the latency is actually better. I saw a benchmark claiming 40ms, but that seems too low... I want to try this tonight. Does anyone know if the API supports batching or just single requests?