Source: MarkTechPost — 2026-07-19
Summary
Feyn AI released SQRL, a family of open text-to-SQL models (4B, 9B, and a 35B mixture-of-experts) that treat query generation as a database-inspection task rather than pure text translation: before committing to a final SQL statement, the model runs read-only probes against the actual database to resolve ambiguity. The flagship SQRL-35B-A3B hits 70.6% execution accuracy on the BIRD Dev benchmark, edging out Claude Opus 4.6's 68.77% on the same evaluation, and even the 4B model matches Opus 4.6's score.
Key Takeaways
- Most text-to-SQL systems guess a query from the question and schema alone; SQRL instead lets the model issue exploratory read-only queries first (checking actual column values, distributions, edge cases) before writing the query it commits to.
- SQRL-35B-A3B: 70.6% BIRD Dev execution accuracy vs. Claude Opus 4.6's 68.77% under the same eval; the 9B model holds nearly all of that at 69.80%, and SQRL-4B reaches 68.80% — matching a frontier closed model in a footprint small enough to self-host.
- All three checkpoints (SQRL-4B, SQRL-9B, SQRL-35B-A3B) are open on Hugging Face, built on the Qwen3.5/Qwen3.6 base families, with model cards including the full system prompt and reference harness.
- The self-hosting angle matters for regulated data: running SQRL locally means schemas, queries, and the model's own inspection probes never leave infrastructure you control — unlike routing every text-to-SQL request through a hosted frontier API.
Reel Script
Hook (17s / 38 words) A small open model just beat Claude Opus 4.6 at writing SQL — not by being smarter, but by doing something obvious that most text-to-SQL systems skip: actually looking at the data before guessing a query.
Core Concept (75s / 165 words) Text-to-SQL is the task of turning a plain-English question into a working SQL query. Most systems do this in one shot: read the question, read the schema — the table and column names — and generate a query, hoping the model's guess about what's actually inside those columns is right. That's where they break, because a schema tells you a column is called "status," not that its real values are "active," "ACTIVE," and "1" all mixed together. Feyn AI's SQRL models fix this by adding a step most people never bothered to build: before writing the final query, the model runs its own small, read-only probe queries against the real database — checking actual value formats, distributions, edge cases — the same way a careful analyst would poke around a spreadsheet before writing a formula. Only after that inspection does it commit to the query it actually returns to you.
Hands-On (100s / 235 words) The benchmark that matters here is BIRD Dev, a standard test set for measuring whether a generated SQL query actually executes and returns the right answer — not whether it merely looks plausible. SQRL-35B-A3B, the largest of the three released checkpoints, scores 70.6% execution accuracy on BIRD Dev. Run the same benchmark against Claude Opus 4.6 and you get 68.77% — so a purpose-built open model edges out a frontier general-purpose one on this specific task. What's more striking is the smaller end: the 9B version holds onto almost all of that accuracy at 69.80%, and even the 4B model — small enough to run on a single consumer GPU — reaches 68.80%, matching Opus 4.6's score outright. All three sizes are published openly on Hugging Face, built on top of the Qwen3.5 and Qwen3.6 model families, and Feyn shipped the full system prompt and reference harness alongside the weights — so you can inspect exactly how the inspect-then-query loop is instructed, not just trust a black-box API to do it right.
Takeaway (25s / 55 words) If your product needs reliable SQL generation over real production data, a small self-hosted model that checks its work before answering is a better bet than a bigger model that doesn't — and it keeps your schema and query logs off someone else's API logs entirely. Worth trying SQRL-4B before reaching for a frontier model on this task.