As billion-dollar bets reshape AI agent infrastructure, a Korean AI startup is offering a free alternative.
TL;DR: VIDRAFT, a Korean AI startup, has positioned its open-weights model Darwin-27B-ZTC-v2 as a direct open-source alternative to Jev, a typed decision model backed by $870 million in venture funding. According to Vietnamese tech outlet Dangmua, VIDRAFT claims Darwin-27B-ZTC-v2 ranks first among 102 models on the System One Mosaic Benchmark decision-engine leaderboard, with a Borda score of 89.58 and a task average of 66.46. The model is released under the Apache-2.0 license, making it freely available for commercial use.
VIDRAFT, a Korean AI startup, has entered one of the most closely watched races in agent infrastructure — the competition to replace general-purpose language models inside agentic decision loops — with a model it says outperforms more than a hundred rivals on a dedicated benchmark. The development comes as TypeSafe AI's Jev, a proprietary typed-decision model founded by ex-OpenAI researcher Diogo Almeida, commands a $7.5 billion valuation after closing an $870 million funding round led by a16z. Where Jev charges for every inference, VIDRAFT is making the opposite bet: that the winning architecture should be open.
VIDRAFT claims its model, Darwin-27B-ZTC-v2, tops the System One Mosaic Benchmark decision-engine leaderboard, finishing first out of 102 evaluated models. The reported scores are a Borda score of 89.58 and a task average of 66.46. The model is released under the Apache-2.0 license, which permits commercial use and modification without royalty obligations.
The approach VIDRAFT describes — referred to as Zero-Token Classification, or ZTC — is conceptually aligned with what Jev does commercially: rather than generating free-form text, the model is designed to return structured outputs suited to decision nodes inside agent graphs. The idea is that most branching logic in an automated pipeline does not require a paragraph-generating language model; it requires a fast, reliable answer to a typed question. VIDRAFT's Darwin-27B-ZTC-v2 is its attempt to prove that a community-accessible model can perform that role as well as, or better than, proprietary alternatives.
The timing is notable. The Dangmua report indicates that Jev only launched on September 15, 2026, and that its debut reportedly accelerated OpenAI's own release of a competing Decisions API. Within that context, VIDRAFT's announcement positions Darwin-27B-ZTC-v2 not merely as a research artifact but as a practical option for developers who want typed-decision capability without vendor lock-in or per-token costs.
The emergence of typed-decision models as a distinct product category is significant because it reflects a maturing view of how agent systems should be architected. The dominant paradigm until recently has been to route nearly everything through a large chat model, even simple binary decisions that could be handled deterministically. Jev's fundraise at a multi-billion-dollar valuation signals that investors believe purpose-built decision models will become a standard layer in production agent stacks — and that the market is large enough to support dedicated companies.
VIDRAFT's move matters for a different reason: it introduces competitive pressure from the open-weights community at exactly the moment when the proprietary market is pricing in enormous future demand. If Darwin-27B-ZTC-v2 can deliver benchmark-leading performance under an Apache-2.0 license, it gives engineering teams a credible reason to avoid committing to a single commercial vendor for a capability that sits at the center of every agent graph.
The Jev architecture, as described in the Dangmua piece, exposes three question types — a multiple-choice selector, an ordered-scale scorer, and a yes/no probability output — and prices access at $0.042 per million input tokens with free output. For high-frequency micro-decisions that would be prohibitively expensive to route through a full language model, that pricing is already competitive. An open-weights alternative changes that calculation further: teams running their own inference infrastructure could potentially bring that marginal cost even lower.
For developers building with frameworks like LangGraph, the practical implication is that a decision node in a state-machine agent graph now has multiple viable backends — proprietary and open — competing on both price and accuracy metrics.
Q: What is VIDRAFT's Darwin-27B-ZTC-v2 and what does it do?
A: Darwin-27B-ZTC-v2 is an open-weights model developed by VIDRAFT, a Korean AI startup, designed to return structured decisions rather than free-form text. It is built for use as a decision node inside AI agent pipelines and is released under the Apache-2.0 license.
Q: How does VIDRAFT's model compare to Jev on the benchmark?
A: According to the Dangmua report, VIDRAFT claims Darwin-27B-ZTC-v2 ranks first among 102 models on the System One Mosaic Benchmark decision-engine leaderboard, with a Borda score of 89.58 and a task average of 66.46.
Q: Why is Jev significant, and how does VIDRAFT fit into that story?
A: TypeSafe AI raised $870 million at a $7.5 billion valuation for Jev, a proprietary typed-decision model aimed at AI agent infrastructure. VIDRAFT positions Darwin-27B-ZTC-v2 as the open-weights alternative to that proprietary approach, offering similar functionality without per-token licensing fees.
Source: 당무아(베트남) (2026-10-12) — original article