Skip to content
← Models
Open weightsAlibaba (XGenerationLab)/10 trustMetrics-checked

XiYan-SQL

Alibaba's QwenCoder-based text-to-SQL family, topping open BIRD and Spider scores in 2025.

Visit official page →

About

XiYan-SQL is Alibaba's XGenerationLab text-to-SQL model series, fine-tuned on the QwenCoder family in 3B, 7B, 14B, and 32B sizes. The line is actively iterated — the v2504 release in April 2025 refreshed the weights — and it is one of the few text-to-SQL models with real product plumbing behind it: Alibaba offers it through ModelScope API inference and an Alibaba Cloud DashScope endpoint alongside the open weights.

The numbers are the story here. The team reports 75.63% execution accuracy on BIRD with an ensemble and 69.03% for the single 32B model, plus 89.65% on Spider, which was state of the art for open models at the time. The weights are Apache 2.0 licensed. For anyone building a text-to-SQL feature in 2025, this is the open baseline to beat.

Specs

Context window

Not published in surfaced materials

Pricing

Free / open weights (Apache 2.0); optional paid API via ModelScope / DashScope

Free quota

Free / open weights — self-host at no license cost

API access

Self-hosted via Hugging Face (XGenerationLab/XiYanSQL-QwenCoder-32B-2502) and ModelScope; optional Alibaba Cloud DashScope API

Weights

Open weights

Released

Not disclosed

Benchmarks

  • BIRD: 75.63% EX ensemble, 69.03% EX single 32B model (provider-reported)
  • Spider: 89.65% EX, claimed SOTA at Nov 2024 (provider-reported)
  • SQL-Eval: 69.86% (provider-reported)

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

text-to-sqldata-analyticsgraph-query-generationbi-copilots

Trust score — 8/10

/10 trustMetrics-checked

Scored 8/10 (metrics tier, metrics-v1). Evidence: Alibaba XGenerationLab releases with versioned model checkpoints (2412, 2502, v2504), Apache-2.0 license, and product-grade API plumbing; benchmark claims are specific and versioned. Strengths: top open text-to-SQL scores, active 2025 maintenance, permissive license. Docked for: (1) all headline numbers are vendor-reported with no independent replication surfaced; (2) the ensemble-vs-single-model distinction is easy to miss when comparing scores.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
8.0/10
Maintenance · recency, releases, issue load
9.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · Versioned checkpoints (2412, 2502, v2504); product API via ModelScope/DashScope
  • · All headline numbers vendor-reported; ensemble-vs-single distinction easy to miss

Sources