A Second Open Challenger to TypeSafe's Jev Emerges, and Hugging Face's Numbers Say Developers Are Trying It
Two models from a little-known Singapore-incorporated lab, AutoTrust AI, posted some of the largest likes gains on Hugging Face in today's ai-tracker snapshot: JEV-27B-VL rose from 883 to 1,489 likes (+68.6%), and its sibling GEV-26B-Decide rose from 629 to 998 (+58.7%). Unlike a typical one-day launch spike that fades, both models have now been climbing for over a week — JEV-27B first shipped September 28-29, with a vision-enabled variant and expanded capabilities added October 1 and 3, and GEV-26B-Decide following close behind on October 3.
What these models actually do
AutoTrust AI's models are purpose-built 'decision models': rather than generating free-form text, they return calibrated probability distributions over a fixed set of options — yes/no questions, multiple-choice with up to 256 options, or 0-to-5 ratings — in a single fast forward pass. JEV-27B pairs a frozen Qwen3.8-27B backbone with a lightweight 108.9-million-parameter trainable decision head (0.4% of the full model), hitting a median decision latency of 137 milliseconds and roughly 130 decisions per second on one Nvidia B200 GPU, while the vision variant JEV-27B-VL extends the same approach to images — reporting 75% success on robot pick-and-place tasks from camera input and 95% completion on multi-step computer-use tasks from screenshots. GEV-26B-Decide applies the identical idea to a Google Gemma-4-26B-A4B backbone instead of Qwen, giving developers a second base-model option.
Jev proved there is real demand for models that decide rather than write. — Daniel Tang, CEO and co-founder, AutoTrust AI
The target is explicit: TypeSafe AI's Jev
AutoTrust's own model cards make the competitive framing unambiguous. JEV-27B's documentation benchmarks itself directly against 'TypeSafe Jev 1.13' by KL divergence on held-out data, and claims a narrow edge on public decision benchmarks (84.07% versus 83.85%). GEV-26B-Decide's card goes further, explicitly stating the model is 'not affiliated with TypeSafe AI' — the kind of disclaimer that only gets written when a resemblance needs disclaiming. TypeSafe's hosted Jev 1.13 model is itself a real, actively used product: it ranked eighth on OpenRouter's token-usage leaderboard in today's report, moving up three places with 3.34 trillion tokens and nearly 982 million requests in a single day, evidence of the exact commercial demand AutoTrust is trying to intercept with a free, self-hostable alternative.
Co-founder Josh Liu framed the pitch around infrastructure economics rather than raw capability: "Every AI agent is really a long chain of small decisions. Make each one fast, private and cheap, and you change the economics," he told RuntimeWire. The argument is that most agentic workflows don't need a frontier model's full generative power for routing, moderation, or classification calls — they need a cheap, fast, locally-hosted judgment call, which is a smaller and more defensible wedge than competing with GPT-6 Astra or Claude Opus 5.5 head-on.
Not the first, and that's the point
AutoTrust isn't alone in this niche. Cloudflare open-sourced its own Jev challenger, Clef, on October 2, and machine.net covered independent quantizers shipping GGUF and MLX builds of Clef within a day of release. AutoTrust's entry, arriving via a company with no prior public track record and a research-tool side product (ScienceGuru) rather than infrastructure pedigree, is a different kind of signal: it suggests TypeSafe's success with Jev has created a genuine, crowded race among open-weights labs to build the cheaper, self-hostable alternative — not a single well-funded competitor, but a pattern. For a tracker that watches download and likes velocity as a leading indicator, two unrelated labs producing sustained (not just launch-day) growth around the same underlying idea within the same week is the more interesting story than either release alone.