Cambridge Enterprise Ventures backs Sqwish Labs' kettei
VenturesCambridge-based Sqwish Labs has secured $3 million in pre-seed funding led by Cambridge Enterprise Ventures, the University of Cambridge investment arm, with Parkwalk and support from Innovate UK, the UK’s innovation agency.
As of 29 October, two weeks after launch, Sqwish Labs’ largest model, kettei fat, ranks first for capability among the hosted decision models on JevBench, an independent benchmark, ahead of those from OpenAI, Microsoft and TypeSafe.
Capability measures how often a model is right and how well its confidence matches reality. On ImageJevBench, which tests decisions about images, kettei fat ranks first overall of 64 systems. It is also second on S1MB’s English leaderboard, ahead of Perplexity and Cloudflare.
Also investing in the round are new angels, including Professor Steve Young, Emeritus Professor of Information Engineering at Cambridge, whose start-up VocalIQ was acquired by Apple for Siri, Karthik Tadinada, CEO of Fortify AI and former COO of Featurespace, and partners from Highland Europe, investing in a personal capacity. Earlier backers, including Vento, Mandeep Singh (former CEO of Trouva), Jon Bradford (partner at Dynamo Ventures) and Guy Fighel (co-founder of SignifAI, acquired by New Relic), also converted their investment into this round.
The latest AI category
When TypeSafe AI launched Jev on 15th September 2026, it brought wider attention to decision models, the specialised, fast AI systems that produce structured data, such as multiple-choice selections, numerical scores, or yes/no probabilities, instead of generating free-form text. In the less than two months since, OpenAI, Microsoft, Amazon, Cloudflare and Perplexity, as well as a handful of well-funded startups, have all launched their own decision models.
Sqwish Labs built its kettei decision models with a tiny fraction of the funding of other similar models, with the company raising $3 million in total funding, in comparison TypeSafe, the company behind Jev, raised $870 million in October. Sqwish Labs’ first products compressed and routed the requests sent to large AI models, and in July its model router took first place on RouterArena, an open leaderboard for model routers. The company shipped its kettei models on 14th October, and ranks first or in the top three on the main independent decision model benchmarks. The company’s models assign a probability to each answer in milliseconds, with no text, and can also read images through its own model variants, kettei fat and kettei chub. Through Sqwish Labs’ OneHot platform, customers can upload examples and get a bespoke model in minutes, running as fast as the original model. Any team can teach kettei its own task in a couple of clicks, and keep teaching it as its work changes, with tests on both the new examples and earlier ones, to check it still does what it learned before.
Decision models will underpin the agentic AI system, whose primary function is to make choices, from which tool to use, to whether to act or whether an answer is safe. As kettei gives each answer a probability, a product can act on the decisions it’s sure of and pass only the hardest cases to a larger model. Prices start at $0.02 per million input tokens: a thousand 500-token decisions on kettei fat cost about three cents.
Two of the models, kettei core and kettei dot, are open-weight and free to download from Hugging Face, so a company can run them on its own computers and keep its data in-house. Teams already using Jev can switch by changing three settings. Any agent can call kettei through Python and TypeScript SDKs, an MCP server or a Claude Code plugin.
Left to right: Dr Ushnish Sengupta and Federica Freddi
System 1 for AI Agents
Sqwish Labs will use the funding to train the next generation of kettei. It goes beyond choosing between options to give numbers with uncertainty, dates and passages of text, in more languages. The company will also scale OneHot for enterprise customers.. It will make its first commercial and developer-relations hires and grow its research team in Cambridge. Innovate UK is co-funding the research behind the kettei models, and the next step: models that keep learning from real-world feedback.
“AI doesn’t need a giant model, or a bigger data centre, for every small decision,” said Ushnish Sengupta, co-founder and CEO of Sqwish Labs. “A small, fast model can make most of them for a fraction of the cost and energy, and knows when to ask for help. This funding takes kettei from the top of the leaderboards into the products people use every day.”
"AI doesn't need a giant model, or a bigger data centre, for every small decision,"
Sqwish Labs was founded in 2024 by Federica Freddi and Ushnish Sengupta, and is headquartered in Cambridge with a headcount of 8. Both University of Cambridge graduates, they met at the institution in 2018, and started building Sqwish in 2024. Sqwish is also in the Slush 100 Top 20, chosen from 1,400+ applicants, and will pitch in the semi-final in Helsinki on 18 November.
Federica Freddi, co-founder and CTO of Sqwish Labs, comments:
“Most AI is frozen the moment it ships. We believe every AI product will run on two systems: a fast “System 1” that makes everyday decisions and learns from what happens after each one, and a large model for the hard cases. At Sqwish, we’re building System 1, so AI products get better, faster and cheaper with every decision they make. We’ve shown that eight people in Cambridge can build one that outranks labs with hundreds of times our funding.”
Alongside the fundraise, Anne Dobrée, Investment Director at Parkwalk, joined the board of Sqwish Labs as investor director. Two board observers sit alongside her, Mahesh Santiapillai, Investment Manager at Cambridge Enterprise, and Laurence Garrett.
“2026 has been the year the AI market grew up. The industry is moving beyond the 'tokenmaxing' mindset toward a far harder standard: proving measurable ROI in live, shifting environments rather than static benchmarks. Sqwish sits at the heart of that shift.”
Mahesh Santiapillai continued, “In Nish and Federica, we saw a deeply technical team building directly against the two problems that will define the next phase of enterprise AI, continuous learning and measurable efficiency.”
“When I first met the team, I was blown away by their technology and its superiority over the competition,” said Laurence Garrett, a Highland Europe partner investing personally. “I was judging an event and simply wrote on my feedback form: ‘How do I invest here?’ Sqwish demonstrates the best of Cambridge and the AI technology scene.”
Anne Dobrée, Investment Director, Parkwalk, adds: “Every AI product makes thousands of decisions a second. Sqwish’s small, specialised models make those decisions faster and more efficiently, and keep improving on live production data. That is a compelling answer to a real commercial problem. Federica and Nish combine deep technical expertise with exceptional speed of execution. When I asked whether kettei could beat my son at Mario Kart, the team built Kettei Kart to find out. That curiosity and pace is exactly what we look for in founders, and Parkwalk is delighted to back them.”