Synthefy is building foundation models for tabular and time series data, the numeric tables that sit behind demand forecasts, pricing decisions, fraud detection, and infrastructure monitoring. Its flagship model, Synthefy-Nori-V1, is an open-source tabular foundation model with 6M parameters that requires zero training and ranks first on mean R² across a 96-dataset regression benchmark . The pitch to data teams is straightforward. Instead of standing up a fresh LightGBM or XGBoost pipeline for every new problem, teams point Nori at a new table and get a working prediction, compressing what the company describes as weeks of prep and tuning into minutes.
The launch matters because tabular ML has largely sat out the foundation model wave, and Synthefy is one of a small group trying to change that. The company was founded in 2023 by CEO Somi Agarwal, previously a motion planning engineer at Uber, and Sandeep Chinchali, an assistant professor at UT Austin , and it ships through three routes: an open model, a managed API, and deployment inside a customer's own environment on AWS, Databricks, Snowflake, or Google Cloud. Early work spans retail, financial services, telecom, infrastructure, healthcare, and defense.
The launch coincides with a $6.5 million seed round led by Wing Venture Capital, with participation from Haystack, Samsung Next, Canonical, and Lightscape , alongside angels Srinivas Narayanan, former CTO of Applied AI at OpenAI, Aparna Chennapragada, Chief Product Officer of Experiences and Devices at Microsoft, and Manohar Paluri, Vice President of AI at Meta Superintelligence Labs . For founders and operators watching where the next model layer gets built, Synthefy is a bet that the reusable pretraining trick that worked for text can work for the numbers that run the physical and financial economy.
Comments (8)
Sign in to join the discussion.
P
Priya VatsanAug 24, 2026
Foundation models for CSVs is a sentence I did not have on my 2025 bingo card, but here we are. Curious what happens when someone points this at a spreadsheet with 47 tabs and merged cells.
T
Tomasz WroblewskiAug 24, 2026
The tweet buries the actual product under the funding announcement. Lead with what it does, then flex the cap table, otherwise half your feed scrolls at the dollar figure.
K
Kenji ItoAug 24, 2026
Genuine question from someone new to this space: how do you evaluate a foundation model for tabular data? Like what does the eval set even look like when every schema is different?
R
Reyhaneh D.Aug 24, 2026
Any docs on the API surface yet? Rate limits, batch inference endpoints, and whether I can fine-tune on private schemas would sell me faster than the seed announcement.
A
Adanna OkekeAug 24, 2026
Rewriting the tagline for free: 'The model that reads your database like GPT reads books.' You can venmo me later.
L
Linus FrombergAug 24, 2026
Before we buy, I need SOC2 Type II, SSO via Okta, a DPA, and a redlined MSA. Also please tell me the data does not leave our VPC.
M
Marco PolveriniAug 24, 2026
Feels wild that in 2025 someone is still shipping a non-agentic product. Give it 8 months and this will be a tool call inside a larger loop, mark my words.
B
Benoit ClementAug 24, 2026
Respectfully, tabular data was solved by XGBoost and we all moved on. What are foundation models doing here that gradient boosting was not already crushing in 2019?
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Synthefy
Foundation models for tabular structured data.
Foundation models for tabular structured data.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Synthefy is building foundation models for tabular and time series data, the numeric tables that sit behind demand forecasts, pricing decisions, fraud detection, and infrastructure monitoring. Its flagship model, Synthefy-Nori-V1, is an open-source tabular foundation model with 6M parameters that requires zero training and ranks first on mean R² across a 96-dataset regression benchmark . The pitch to data teams is straightforward. Instead of standing up a fresh LightGBM or XGBoost pipeline for every new problem, teams point Nori at a new table and get a working prediction, compressing what the company describes as weeks of prep and tuning into minutes. The launch matters because tabular ML has largely sat out the foundation model wave, and Synthefy is one of a small group trying to change that. The company was founded in 2023 by CEO Somi Agarwal, previously a motion planning engineer at Uber, and Sandeep Chinchali, an assistant professor at UT Austin , and it ships through three routes: an open model, a managed API, and deployment inside a customer's own environment on AWS, Databricks, Snowflake, or Google Cloud. Early work spans retail, financial services, telecom, infrastructure, healthcare, and defense. The launch coincides with a $6.5 million seed round led by Wing Venture Capital, with participation from Haystack, Samsung Next, Canonical, and Lightscape , alongside angels Srinivas Narayanan, former CTO of Applied AI at OpenAI, Aparna Chennapragada, Chief Product Officer of Experiences and Devices at Microsoft, and Manohar Paluri, Vice President of AI at Meta Superintelligence Labs . For founders and operators watching where the next model layer gets built, Synthefy is a bet that the reusable pretraining trick that worked for text can work for the numbers that run the physical and financial economy.
Comments (8)
Foundation models for CSVs is a sentence I did not have on my 2025 bingo card, but here we are. Curious what happens when someone points this at a spreadsheet with 47 tabs and merged cells.
The tweet buries the actual product under the funding announcement. Lead with what it does, then flex the cap table, otherwise half your feed scrolls at the dollar figure.
Genuine question from someone new to this space: how do you evaluate a foundation model for tabular data? Like what does the eval set even look like when every schema is different?
Any docs on the API surface yet? Rate limits, batch inference endpoints, and whether I can fine-tune on private schemas would sell me faster than the seed announcement.
Rewriting the tagline for free: 'The model that reads your database like GPT reads books.' You can venmo me later.
Before we buy, I need SOC2 Type II, SSO via Okta, a DPA, and a redlined MSA. Also please tell me the data does not leave our VPC.
Feels wild that in 2025 someone is still shipping a non-agentic product. Give it 8 months and this will be a tool call inside a larger loop, mark my words.
Respectfully, tabular data was solved by XGBoost and we all moved on. What are foundation models doing here that gradient boosting was not already crushing in 2019?