Text-to-SQL has a confidence problem. Most models can write syntactically flawless SQL that executes without a hiccup—and still returns the wrong numbers. The issue, as many enterprise teams have discovered the hard way, isn’t that the model doesn’t understand SQL; it’s that the schema alone can’t tell you whether a county is stored as "Alameda," "Alameda County," or "ALAMEDA." Without peeking at the actual data, the model is guessing.

Feyn AI, a Y Combinator–backed startup, is betting that the fix isn’t a bigger model or fancier prompts—it’s giving the model permission to look. On July 19, 2026, the company released SQRL, a family of open-weight text-to-SQL models that take an inspection-first approach: instead of translating a question directly into a query, the model can fire off read-only probes into the database, see what comes back, and only then commit to a final SQL statement.

The flagship model, SQRL-35B-A3B, a mixture-of-experts architecture with about 3 billion active parameters out of a total 35 billion, achieved 70.6% execution accuracy on the BIRD Dev benchmark. Under the same evaluation harness, Claude Opus 4.6 scored 68.77%. Both numbers underscore how brutal BIRD is—it uses 95 real databases spanning 37 professional domains and 33.4 GB of actual data, with execution-verified grading. In other words, it’s a lot closer to production than academic benchmarks like Spider, where GPT-4o can hit 86.6% but crater to 10.1% on the more realistic Spider 2.0.

Inspect First, Write Later

SQRL’s architecture is built around two output actions. When the model receives a question and the database schema, it can either emit a <sql> block to request a read-only observation—something like SELECT DISTINCT CountyName FROM schools LIMIT 10—or it can emit an <answer> block with the final query. The harness executes exploration queries in a sandboxed, read-only mode and returns the rows inside <observation> tags. The model gets up to five inspection steps, though most questions get sorted out in one or two.

This isn’t just a clever prompting trick. The model was trained to decide when probing is worth the extra latency. For straightforward counts or unambiguous schemas, SQRL answers immediately. When column semantics are murky or the data has surprise formats, it probes. "SQRL combines both strategies in one model: cheap single-shot behavior for easy questions, and conditional inspection for ambiguous ones," noted developer commentary site Kiadev. "The system only pays the runtime cost of probes when needed."

Three Sizes, One Training Pipeline

The SQRL family comes in three flavors:

The teacher model (35B-A3B) was trained using CISPO, a clipped importance sampling policy optimization algorithm that keeps reinforcement learning stable at scale. For each training question, the teacher generated eight trajectories, and the team provided a binary reward when the executed SQL matched the reference answer. They specifically trained on the "mixed zone"—cases where some trajectories succeeded and others failed—so the model would learn which choices actually drove success. Around 10,200 verified multi-step trajectories were then used to distill the smaller 4B and 9B models.

A Crowded Field with a Clear Differentiator

Text-to-SQL has become one of the most heavily contested corners of enterprise AI. Snowflake sells Cortex Analyst as a fully managed, semantic-model-powered service that runs inside its platform. Databricks counters with Genie Spaces, baked into its Lakehouse. Both promise to handle governance and scale natively, and both lean heavily on platform lock-in.

SQRL’s bet is different: open weights, self-hosted via vLLM, with an architecture that doesn’t just translate—it investigates. "The missing information already lives inside the database," MarkTechPost observed. "The model simply needs permission to ask for it." That permission, of course, comes with trade-offs. Read-only probes mean extra queries hitting the database, potentially raising data access audit flags. In environments governed by regulations like GDPR or China’s Personal Information Protection Law, even read-only access must respect the principle of minimal necessity. Teams deploying SQRL will need to log and review those probes just as carefully as any production SQL.

On raw accuracy, the field is moving fast. In June 2026, Google Research unveiled Gemini-SQL2, a Gemini 3.1 Pro-based system that hit 80.04% on BIRD, setting a single-model record. That’s a different harness and likely a different post-training regimen, but it shows the ceiling isn’t static. Still, as one dbt Labs benchmark put it, "Text-to-SQL will cheerfully give you a wrong number," and the gap between benchmark scores and production reality remains stark. Cost can also be a pragmatic factor: running an 8×H200 node for self-hosted inference can run $30–50 per hour, and the break-even point over managed APIs tends to hover around 3,000 prompts per day.

Early Chatter and Open Questions

The release is too fresh for a mature community verdict—the model pages on Hugging Face are just starting to accumulate downloads—but the tone so far has been intrigued rather than dismissive. Byteiota, covering the launch, wrote, "This is not a marginal improvement. It is a different architecture." IT analyst site itinai.com added, "By letting the model look at the data before it writes SQL, practitioners eliminate the class of wrong-but-valid queries that have plagued text-to-SQL systems for years."

Some developers, however, remain wary. Letting a model issue arbitrary read-only queries feels, to some DBAs, like handing a junior analyst a read-only console and hoping for the best. The probes are explainable in retrospect—you can see what the model asked and why—but the decision to probe is still a black box at runtime. Whether inspection-first architectures can earn the kind of trust that lets them sit inside a financial reporting pipeline or a healthcare analytics stack is the real test.

For now, SQRL gives teams a new lever. In a market that QYResearch estimates will grow from $47.4 billion in 2025 to over $164 billion by 2032, and where Gartner reports that more than 60% of large data teams are already piloting text-to-SQL, the ability to self-host an open model that can literally double-check its own work might be exactly what the enterprise has been waiting for. Or, as one early observer put it, "The DBA of 2026 is less a query writer and more a system architect, semantic modeller, and AI governor." SQRL is a tool for that new reality.