Sqwish Labs Raises $3M Pre-Seed for AI Decision Models
Two words at the end of a legal definition can change the answer a reviewer needs. In a merger-agreement experiment published by Sqwish Labs, the phrase “after due inquiry” changed how a knowledge provision should be classified. That small piece of language gives the Cambridge AI company's financing a practical destination: helping software learn the distinctions its users already have to make.
Sqwish Labs has secured a $3M Pre-Seed led by Cambridge Enterprise Ventures, with Parkwalk participating and Innovate UK supporting research. The Cambridge Enterprise announcement, published November 18, 2026, identifies co-founder and CEO Ushnish Sengupta and co-founder and CTO Federica Freddi. The financing backs development of the kettei decision-model family, enterprise expansion of OneHot and hiring across commercial, developer-relations and research work.
For teams building AI products, the proposition puts a familiar operating responsibility inside the model purchase. A customer defines the choices, supplies examples and decides which mistakes deserve intervention. Sqwish's opportunity lies in making those judgments usable by smaller models, while keeping the person responsible for the workflow connected to how those models change.
What the $3M Pre-Seed supports
Cambridge Enterprise Ventures leads the financing alongside Parkwalk, a UK university-technology investor. New individual investors include Steve Young and Karthik Tadinada, with Highland Europe partners participating personally. The announcement also says earlier investments converted into the round. Its $3M total-funding figure therefore includes that earlier backing; adding it again would overstate the capital available to the company.
Innovate UK's involvement has a separate role in the story. The agency co-funds research behind kettei and the company's work on models learning from feedback. The announcement does not disclose a detailed split between investment and research support, individual check sizes or a valuation. It presents the amount in dollars without explicitly naming the denomination, so the reported $3M is retained without a currency conversion.
The operating priorities reach beyond another model release. Sqwish plans to broaden kettei's capabilities, expand OneHot for enterprise customers and bring in its first commercial and developer-relations staff. That hiring agenda puts a customer-facing assignment beside the research: understand which decisions a buyer wants to delegate and help the buyer express them clearly enough for the product to serve them.
The company reports a Cambridge team of 8, founded in 2024. That makes the first commercial hires a meaningful change in how the business meets its market. Engineers can publish an API and a benchmark; enterprise adoption brings conversations about examples, integration and responsibility for the results back into the room.
How kettei gives a buyer control over the question
Sqwish's October 14 kettei launch describes a model that receives context, a question and permitted answers, then scores those answers with probabilities. A product developer can use that output to choose a tool, classify a request or escalate a case. The surrounding application still determines what actions are permitted and how much confidence is enough to proceed.
The product family gives developers different deployment choices. The larger fat and chub variants accept text and images, while core and dot are smaller text models with downloadable weights. OneHot provides hosted inference and fine-tuning. The current model page lists core at $0.02 per million input tokens, with higher rates for chub and fat, making model selection part of the buyer's cost calculation.
A useful probability changes the design of an escalation rule. The application can reserve a larger model or human review for cases falling below its threshold, while allowing routine choices to move through a faster path. That is an architectural possibility, rather than a measured saving promised to every customer. Related continual-learning work at Remanence puts the feedback question elsewhere in the AI infrastructure market. Its usefulness depends on the task, the reliability of the scores and the consequences of a wrong choice.
Sqwish's own release acknowledges that frontier language models score higher on some comparisons while taking longer. This leaves room for a purchasing conversation about the right model for each job. A team handling repeated, bounded choices can evaluate specialized inference without assuming the same model should also write its customer response or reason through an unfamiliar assignment.
What the contract experiment says about customer knowledge
The November 13 fine-tuning report trained kettei chub on 122 merger agreements and evaluated it on 30 unseen agreements. Sqwish reports accuracy improving from 86.5% to 94.5%. These are results from the company's experiment, with a disclosed task and dataset, rather than a measurement of a named customer's production workload.
The legal example explains why the training material matters. A phrase containing the words “actual knowledge” could still receive a different classification because an obligation to inquire followed it. A customer buying faster document review has to preserve precisely that sort of distinction. Teaching a model requires access to judgments that someone understands well enough to label, inspect and correct.
Sqwish's report also describes a feedback process that selects useful new cases, retains earlier examples and tests candidate updates before replacement. It discusses checking feedback for contradictory labels and missing context. That gives continual learning an operational meaning: a change in model behavior needs a reason, supporting examples and a way to establish that earlier useful behavior survived.
The founders' account of Sqwish's origins describes their earlier product work being consumed by prompt tuning, context management and model choices. Their move into infrastructure follows that experience. The new commercial and developer-relations hires will now encounter other teams' versions of the same labor, with different documents, queues and definitions of an acceptable answer arriving alongside the request for a faster model.
AI Infrastructure funding, last 30 days
DevCuration's funding database tracked 34 AI Infrastructure rounds totaling $30.3B in disclosed capital over the past 30 days. Recent deals we covered:
- Linq Raises $90M Series B, Adds Ben Horowitz to BoardSeries B · $90M · Nov 18
- Compute Desk Raises $45M Seed for GPU Compute MarketsSeed · $45M · Nov 18
- BXDC Agrees to $5B AWS Data Center Portfolio Acquisition$5B · Nov 17
- Daisytuner Raises €6M Seed to Expand Daisyflow AI InferenceSeed · €6M · Nov 16
- Zeromatter Discloses $160M Raised at $1.1B Valuation$160M · Nov 16
Frequently Asked Questions
How do decision models change the way an AI application handles routine work?
Kettei scores a defined set of possible answers. An application can use those probabilities to choose an action or send an uncertain case for additional review, with the application owner setting the decision rules.
What does a company need to fine-tune kettei for its own task?
A customer needs examples of the decisions it wants the model to make and reliable answers for those examples. The quality and relevance of those judgments affect what the tuned model can learn.
How should buyers interpret the merger-agreement results?
Sqwish reports a company-run experiment using 122 training agreements and 30 unseen test agreements, with accuracy rising from 86.5% to 94.5%. Buyers should evaluate their own documents and consequences rather than assume the same result transfers to every setting.
How does Innovate UK support differ from the round investors?
Cambridge Enterprise identifies Innovate UK as a research co-funder, while Cambridge Enterprise Ventures leads the Pre-Seed with Parkwalk participating. The announcement does not provide a detailed equity-versus-research-support allocation.
Where the Money Moved
The intelligence briefing of the innovation economy. Funding, M&A, debt and fund closes, read as market signal rather than deal announcements.
Subscribe to Where the Money Moved