Defog Review (2026): Features, Pricing & Verdict
Defog (defog.ai) is a Y Combinator–backed company best known for SQLCoder — a family of open-source large language models purpose-built to convert natural-language questions into SQL, and one of the genuine open-source success stories in text-to-SQL. The pitch that put Defog on the map was concrete and surprising: a comparatively small, specialised model beating giant general-purpose ones at the narrow job of SQL generation. SQLCoder models (spanning 7B, 15B, 34B and 70B parameters in the original run, with later 8B/14B/32B releases) are fine-tuned on base models like StarCoder and CodeLlama over 20,000+ human-curated questions across multiple schemas, and on Defog’s sql-eval framework they outperform GPT-4 and GPT-4-turbo on text-to-SQL while significantly beating all popular open-source models — SQLCoder-7B famously landed better than a fresh GPT-4 on SQL generation out of the gate, and racked up 100k+ Hugging Face downloads. The weights are openly available (CC BY-SA 4.0, commercial use permitted), which means privacy-critical organisations can run state-of-the-art text-to-SQL entirely on their own hardware — SQLCoder-7B fits on a consumer GPU like an RTX 4090 or an Apple M2 Pro, larger variants on A100-class cards — with no per-query API fees and no data ever leaving their servers. That self-hosting story is Defog’s reason for being: the models were built explicitly because enterprise customers in healthcare, finance and government refused to send sensitive data or schemas to third-party APIs. On top of the models Defog built an enterprise platform — a privacy-centric product that connects to your database, fine-tunes on your metadata (Defog claims 99%+ accuracy when fine-tuned), auto-visualises results, and orchestrates SQL/Python/R workflows with human-in-the-loop agents. The important 2026 caveat: Defog is a tiny (~7-person), bootstrapped team that received an M&A offer in April 2025, so while SQLCoder remains freely available and widely used, the standalone commercial platform’s long-term direction is genuinely uncertain.
- Best for
- Privacy-critical orgs (healthcare, finance, government) needing self-hosted, on-prem text-to-SQL where data and schemas cannot leave their infrastructure
- Platform
- SQLCoder open models (Hugging Face, self-hosted); enterprise platform connects to SQL databases/warehouses; cloud or on-premises deployment; read-only focus
- Key differentiator
- State-of-the-art open-source text-to-SQL models that beat GPT-4 on sql-eval, run on your own hardware — no data leaves your servers, no API fees
- Pricing
- SQLCoder weights free (CC BY-SA 4.0, commercial use OK); enterprise cloud/self-hosted tiers enterprise-priced (~$5,000/mo cited for unlimited cloud)
- Vendor
- Defog.ai (YC W23) — founded 2023 by Rishabh Srivastava & Medha Basu; ~$2.7M raised, ~$1.1M revenue, 7-person team; received M&A offer April 2025
What Is Defog?
Defog exists because of a specific enterprise objection: the organisations that most need to democratise data access — hospitals, banks, government agencies — are precisely the ones that cannot send their data, or even their database schemas, to a third-party cloud API. That rules out most text-to-SQL tools, which route your question and metadata to OpenAI or Anthropic. Defog’s founders recognised that the answer was a self-hosted model good enough to rival the closed-source giants, and built SQLCoder to fill that gap. It’s a family of LLMs fine-tuned specifically for one job — turning natural language into SQL — and the specialisation pays off: on Defog’s own sql-eval benchmark, SQLCoder outperforms GPT-4 and GPT-4-turbo at SQL generation despite being a fraction of their size, embodying the founders’ thesis that for narrow tasks, smaller models are better, cheaper and faster. The models are genuinely open (CC BY-SA 4.0 weights, Apache-2 code), downloadable from Hugging Face, and runnable on hardware a single team can afford, so a regulated organisation can stand up state-of-the-art text-to-SQL entirely inside its own firewall. Around the models Defog built a commercial platform for enterprises that want more than raw weights: it connects to your SQL database or warehouse, fine-tunes on your specific metadata for higher accuracy, converts results into tables and charts automatically, and coordinates multi-step SQL/Python/R analytical workflows through human-in-the-loop agents (its “Introspect” line is framed as deep research for internal data). Within this site’s Data Analysis, BI & Spreadsheets category, Defog occupies a distinctive spot: less a polished SaaS platform than a model-first, privacy-first foundation — the choice when self-hosting and data residency outrank turnkey convenience, and where the open SQLCoder weights are often the real product teams adopt.
Core Features
SQLCoder: open-source, GPT-4-beating text-to-SQL models
SQLCoder is Defog’s flagship and its genuine claim to fame. It’s a family of state-of-the-art open LLMs for natural-language-to-SQL, fine-tuned from strong code bases (StarCoder for the 15B, CodeLlama for the 70B) on a hand-curated dataset of 20,000+ questions spanning multiple database schemas — with the crucial methodological rigour that none of the evaluation schemas appeared in training, so benchmark results reflect real generalisation rather than memorisation. The performance is the headline: on Defog’s sql-eval framework, SQLCoder outperforms GPT-4 and GPT-4-turbo at SQL generation and significantly beats all popular open-source models, with even the 7B variant matching or exceeding a fresh GPT-4 on SQL tasks and the 70B reaching around 96% on standard benchmarks after Defog pushed accuracy up roughly 30 percentage points over five months of iteration. Practical accessibility is a major strength: the weights are openly licensed (CC BY-SA 4.0, commercial use allowed, with the caveat that fine-tuned derivatives must be open-sourced under the same terms), available on Hugging Face with 100k+ downloads, and sized for real hardware — SQLCoder-7B runs on a consumer RTX 4090 or Apple M2 Pro with 20GB+ memory, SQLCoder-34B on a 4×A10 or A100-class setup. That means no per-query API fees and full self-hosting. Two honest limits: the models are designed for read-only database interactions and are explicitly unsuited to write operations (INSERT/UPDATE/DELETE), and like all text-to-SQL they degrade on databases with hundreds of tables or messy naming — schema-aware context (metadata pruning, RAG) matters, and complex analytical queries may need iteration.
The enterprise platform: fine-tuning, visualisation and agents
Beyond the open weights, Defog offers a commercial platform for enterprises that want a managed, higher-accuracy, workflow-capable product rather than raw models to operate themselves. Its defining feature is metadata fine-tuning: Defog fine-tunes models on your specific enterprise schema and business context, and claims this pushes accuracy above 99% for text-to-SQL on your data — a meaningful jump over generic model performance, and the kind of accuracy regulated analytics demands. The platform connects to any SQL database or data warehouse with a privacy-centric posture (Defog’s stated position is that your data is never shared with anyone, including its own AI model), lets business teams ask questions in natural language and get quick results, and automatically converts outputs into tables and charts. It reaches past single-query text-to-SQL into workflow automation, managing complex multi-step SQL, Python and R analyses with AI assistance, and uses human-in-the-loop agents to orchestrate task execution and let users edit and collaborate on generated reports — the “Introspect” capability positioned as deep research across internal data. Deployment flexes to privacy requirements: cloud or fully on-premises, so sensitive-data organisations keep everything in their environment. The honest caveats are about maturity and scale rather than capability: Defog is a very small team, independent user reviews of the commercial platform are sparse compared with its widely-benchmarked open models, and the 99%-when-fine-tuned figure is the vendor’s own claim, so it warrants a proof-of-concept on your real schema. The platform’s real-world proof points concentrate in healthcare, finance and government — exactly the privacy-bound sectors it was built for.
Privacy, self-hosting and the small-team reality
Defog’s entire value proposition rests on privacy and control, and that’s where it’s strongest — and where its risks live too. The self-hosting story is best-in-class for regulated buyers: open SQLCoder weights you can run inside your own firewall on hardware you already own, so database schemas and query data never touch an external API, no per-query fees accrue, and you retain complete control over the model and its outputs. For a hospital, bank or government agency that has ruled out cloud text-to-SQL on compliance grounds, Defog (alongside self-hosted options like Vanna) is one of a very small number of viable choices, and that scarcity is much of its appeal. The enterprise platform extends the same philosophy with on-premises deployment and a no-data-sharing stance. But the 2026 reality demands clear eyes about vendor risk. Defog is a bootstrapped, roughly 7-person company with about $1.1M in revenue and ~$2.7M raised, and it received an M&A offer in April 2025 — so while the SQLCoder open models are safely available regardless of corporate outcomes (that’s the beauty of open weights), the standalone commercial platform’s roadmap, support and long-term future are genuinely uncertain. The pragmatic read: the open models are a low-risk adoption — they exist, they’re free, they work, and no acquisition can take them away. The commercial platform is a higher-risk bet on a tiny vendor in flux, so evaluate it with contingency planning, and weight the open-weights path heavily if long-term vendor stability matters to you.
Scored Categories
Pricing
| Offering | Price | Notes |
|---|---|---|
| SQLCoder open models | Free | Weights on Hugging Face (CC BY-SA 4.0); commercial use permitted; self-host on your own hardware — no API fees |
| Self-hosting cost | Your hardware | SQLCoder-7B on RTX 4090 / Apple M2 Pro (20GB+); 34B on 4×A10 or A100-class — you supply the compute |
| Enterprise cloud | ~$5,000/month (cited) | Third-party-cited figure for unlimited cloud usage; fine-tuning, visualisation, agents — confirm with Defog |
| Enterprise self-hosted | Custom | On-premises deployment, metadata fine-tuning, workflow agents — quote-based for regulated organisations |
Strengths
- SQLCoder beats GPT-4 and GPT-4-turbo on Defog’s sql-eval text-to-SQL benchmark
- Fully open weights (CC BY-SA 4.0) — free, commercial use, no lock-in
- Best-in-class privacy: self-host so data and schemas never leave your servers
- Runs on accessible hardware — 7B on a consumer RTX 4090 or Apple M2 Pro
- No per-query API fees when self-hosted
- Rigorous eval methodology — test schemas excluded from training data
- Enterprise platform fine-tunes on your metadata (99%+ accuracy claimed)
- Auto-visualisation plus SQL/Python/R workflow agents with human-in-the-loop
- One of few viable choices for healthcare, finance and government data residency
Weaknesses
- Very small (~7-person), bootstrapped team — limited support capacity
- Received M&A offer April 2025 — commercial platform future uncertain
- Raw models are developer tools, not turnkey for business users
- Read-only focus — not designed for INSERT/UPDATE/DELETE operations
- Self-hosting requires GPU hardware and ML/infrastructure skills
- Commercial pricing opaque; 99%-when-fine-tuned is a vendor claim
- Degrades on databases with hundreds of tables or messy naming
- Sparse independent reviews of the commercial platform vs the open models
Verdict: 7.4 / 10 — Exceptional Open Models and a Best-in-Class Privacy Story, Wrapped in Vendor Uncertainty
Defog scores 7.4 on the strength of a genuinely outstanding open-source contribution: SQLCoder is a state-of-the-art text-to-SQL model family that beats GPT-4 on SQL generation, runs on affordable hardware, and — because it’s fully open and self-hostable — solves the data-residency problem that locks regulated organisations out of every cloud text-to-SQL tool. For healthcare, finance and government teams that cannot send data or schemas to third-party APIs, that combination is close to unique and highly valuable, and the open weights are a low-risk adoption no corporate outcome can undo. The score is pulled down by the commercial side: a tiny bootstrapped vendor that received an M&A offer in April 2025, opaque enterprise pricing, sparse independent reviews of the platform, and raw models that demand developer skills. The pragmatic verdict: adopt the open SQLCoder models with confidence; approach the commercial platform with contingency planning and a proof-of-concept first.
Frequently Asked Questions
Is Defog’s SQLCoder really better than GPT-4 at text-to-SQL?
On the specific, narrow task of SQL generation, Defog’s benchmarks say yes — and the methodology is more rigorous than most vendor claims. On Defog’s sql-eval framework, SQLCoder models outperform GPT-4 and GPT-4-turbo at converting natural language to SQL, with even the small 7B variant matching or beating a fresh GPT-4 on SQL tasks and the 70B reaching around 96% on standard benchmarks. Crucially, Defog excluded all evaluation schemas from the training data, so the results reflect genuine generalisation to unseen databases rather than memorisation — a discipline many “we beat GPT-4” claims skip. The important context: this is task-specific superiority, not general superiority. SQLCoder is a specialist fine-tuned for one job; GPT-4 is a generalist. The founders’ own framing captures it — for narrow tasks, smaller specialised models are often better, cheaper and faster, while for general work the big models win. So the honest answer is that SQLCoder genuinely beats GPT-4 at SQL generation on Defog’s benchmark, which is a real and useful result, but you should still validate on your own schema (benchmark performance and your-database performance can differ), and remember that the practical win isn’t only accuracy — it’s that SQLCoder delivers that accuracy while running on your own hardware with your data never leaving your servers, which GPT-4 via API cannot.
What hardware do I need to self-host SQLCoder?
Less than you might expect, which is much of SQLCoder’s appeal. The model comes in several sizes tuned to different hardware. SQLCoder-7B is the accessible entry point: it runs on a consumer GPU like an NVIDIA RTX 4090, or an Apple M2 Pro/Max/Ultra with 20GB+ of memory — hardware a small team or even an individual can own, with CPU inference possible but slower. Larger, more accurate variants need more: SQLCoder-34B calls for roughly a 4×A10 setup or a consumer GPU with 20GB+ VRAM, and the original testing was done on an A100 40GB with bfloat16 weights. The practical implication is significant for privacy-critical organisations: you can stand up state-of-the-art text-to-SQL entirely inside your own firewall without renting cloud GPUs or paying per-query API fees, and you can start small (7B on existing hardware) to validate the approach before investing in bigger models or more compute for higher accuracy. You’ll need ML/infrastructure skills to deploy and serve the model — this is a developer undertaking, not a click-to-install product — but the hardware bar is genuinely accessible, which is exactly why SQLCoder has 100k+ Hugging Face downloads and is a go-to for teams that need on-prem text-to-SQL.
Should I worry about Defog’s small size and the 2026 M&A offer?
It depends entirely on which part of Defog you’re adopting — and this distinction is the most important thing to understand before choosing it. If you’re using the open SQLCoder models, the vendor’s corporate situation is largely irrelevant to you: the weights are openly licensed (CC BY-SA 4.0), already downloaded 100k+ times, and permanently available regardless of what happens to the company. That’s the enduring value of open weights — no acquisition, pivot or shutdown can retract a model that’s already public and running on your hardware. So for the open-models path, Defog’s small size and the April 2025 M&A offer are essentially non-issues; you own your deployment. If, however, you’re evaluating the commercial enterprise platform — the fine-tuning service, hosted deployment, workflow agents and ongoing support — then vendor stability genuinely matters, and you should proceed with clear-eyed contingency planning. Defog is a roughly 7-person bootstrapped company (~$1.1M revenue, ~$2.7M raised) that received an acquisition offer, so the platform’s roadmap, pricing and support continuity carry real uncertainty. The pragmatic approach: treat the open models as a safe, low-risk foundation you can build on confidently, and treat the commercial platform as a higher-risk bet requiring a direct conversation about current status, support commitments and what happens to your deployment under various corporate outcomes.