The engine claimed a top score on the BIRD Mini-Dev benchmark — without fine-tuning or labeled SQL examples.
TL;DR: VIDRAFT, a Korean AI startup, publicly demonstrated ViSQL, an NL2SQL engine that converts natural-language questions into SQL queries and validates the results before returning them as structured reports. Running on VIDRAFT's Darwin-180B-RSI language model, ViSQL scored 83.80% execution accuracy on the BIRD Mini-Dev benchmark (SQLite, 500 questions), surpassing the current leaderboard leader's 81.00 points. VIDRAFT has submitted the results for official listing and plans to attempt the full BIRD evaluation next.
Korean AI startup VIDRAFT unveiled ViSQL (pronounced "Vi-SQL") on October 30, 2026 — an NL2SQL engine designed to let anyone query a database using everyday language and receive an automatically verified, report-ready answer in return. The public demonstration marked the company's first major product reveal in the data-analytics tooling space and arrived alongside a benchmark result that VIDRAFT says beats every publicly listed score on the BIRD Mini-Dev leaderboard.
ViSQL takes a natural-language question — typed in Korean or English — and converts it into a SQL query, executes that query against a connected database, checks the result for correctness, and then formats everything into tables, charts, or a downloadable DOCX report. Crucially, the engine does not require prior model fine-tuning on task-specific SQL examples; instead, it verifies answers by actually running candidate queries against a live (read-only) database and comparing outcomes.
The engine is built on top of Darwin-180B-RSI, VIDRAFT's flagship large language model. The combined system — referred to internally as Darwin-180B-NL2SQL — achieved an execution accuracy of 83.80% on the BIRD Mini-Dev benchmark (SQLite edition, 500 questions, 2026 revised version), correctly answering 419 of 500 questions. The current top score on the Mini-Dev leaderboard stands at 81.00, making VIDRAFT's self-reported result 2.8 percentage points higher. For context, BIRD's creators — researchers from the University of Hong Kong and Alibaba, who introduced the benchmark at NeurIPS 2023 — peg human expert (data-engineer) performance at roughly 93 points. VIDRAFT has submitted its predictions and scoring artifacts to the BIRD team for independent verification and official listing.
The public demo is live and supports four data scenarios:
Safety guardrails built into the engine include SQL syntax validation, read-only database connections, query execution time limits, and audit logging.
The NL2SQL problem — getting AI to translate human intent into precise, executable database queries — has been one of the more stubborn challenges in applied AI. Errors compound quickly: a subtly wrong SQL clause can silently return incorrect numbers that propagate into decisions. VIDRAFT's approach of having the engine verify its own output by executing queries and inspecting results rather than just predicting syntax is a meaningful architectural choice, and claiming the top BIRD Mini-Dev score without labeled training SQL is a notable claim if independently confirmed.
The two focus verticals — financial compliance (fraud tracing) and public-sector analytics — are areas where data-query errors carry real regulatory and operational risk, making verified output especially valuable. A company representative noted, as reported by 조선비즈, that the most time-consuming part of real-world data work is not asking the question but the subsequent steps of matching standards and locating the right data — a gap ViSQL is explicitly designed to close.
Darwin-180B-RSI, the model underpinning ViSQL, already holds Hugging Face-certified top rankings on math benchmarks including AIME 2026 and HMMT 2026 (perfect scores), science reasoning (GPQA Diamond: 94.44), and document extraction (ExtractBench: 90.29). Across 53 recognized public benchmarks, VIDRAFT claims first-place finishes on 15 — the most of any organization worldwide.
Q: What is ViSQL and who made it?
A: ViSQL is an NL2SQL engine developed by VIDRAFT, a Korean AI startup, that converts plain-language questions into SQL queries, executes them against a database, validates the results, and returns formatted reports — all without requiring fine-tuned SQL training data.
Q: How did ViSQL perform on the BIRD benchmark?
A: On the BIRD Mini-Dev benchmark (SQLite, 500 questions, 2026 edition), ViSQL achieved 83.80% execution accuracy — correctly answering 419 questions — which VIDRAFT reports is 2.8 percentage points above the current top score of 81.00 on the official leaderboard. VIDRAFT has applied for official listing.
Q: What industries is ViSQL targeting first?
A: VIDRAFT is positioning ViSQL for finance and public-sector use cases, with demo scenarios covering anti-money-laundering voice-phishing fund tracing and government budget management, alongside drug-discovery data queries and general user-uploaded datasets.
Source: 조선비즈 (2026-10-30) — original article