Krea is handling large-scale creative production, including a campaign that went from signup to broadcast in two days.
2
Featherless AI argues that reliable agents for ordinary tasks are more valuable than models that are slightly smarter at difficult reasoning.
3
OpenRouter gives developers one API for comparing, routing, and operating across language models.
Summary
This Demo Day features startup pitches from the HF0 residency. Krea presents an AI creative suite already used to generate millions of images for a studio and to produce a Fox advertising broadcast within two days of signup. OpenHome is building an open, developer-oriented smart speaker platform with free dev kits. Coframe wants websites to adapt to each visitor and gives customer experiences an AI growth team. Featherless AI focuses on model reliability, claiming that agents need to complete tasks such as booking flights, handling email, and filing invoices consistently. Upside turns messy sales and marketing records into structured data and a map of business interactions. OpenAudio pitches controllable, expressive voice generation through its S1 model. OpenRouter provides a marketplace and unified API for language models, with routing, privacy controls, observability, rankings, and model-specific features.
Krea is built for the traffic created by massive AI content production
Diego Rodriguez compares predictable changes such as replacing a horse with an engine to the harder problem of traffic. With AI, the difficult questions appear after generation: how do you find one item among a million images produced each day for a studio, or manage personalized content made for different towns in India? Krea was already used by Fox to broadcast an advertisement to millions of people. Diego says the customer went from signup to conversion, payment, and broadcasting in two days. He also presents Krea's 25 million users, its fundraising, and a team of eight.
OpenHome wants developers to build natural voice interfaces in any form
OpenHome argues that smart speakers remain limited because people can barely talk to them and cannot customize them or build communities around them. Its proposed interface uses natural conversation instead of awkward command language. The company is opening the platform to developers, who can create custom speakers, voice applications, games, personalities, and home automation tools. The hardware can take many forms, including talking toys, robots, and appliances. OpenHome says more than 10,000 developers are building on the platform, and it offers 500 free dev kits at the event.
Josh says the internet is static, impersonal, and built around one-size-fits-all websites. Coframe gives each customer experience its own AI growth team, with the goal of making websites adapt to visitors. He cites work that generated $20 million for Europe's largest travel company in a few weeks and increased clickthrough rate for India's largest company, a $400 billion enterprise, over a similar period. Coframe also says it is the only marketing technology company partnered directly with OpenAI at the time of the pitch.
Featherless AI puts agent reliability ahead of model intelligence
Eugene presents Featherless AI as a research lab focused on reliable personalized AGI. He says his team built Quorki 72B, a large model without transformer attention, using eight GPUs, and claims its architecture enables much lower inference costs while matching performance. His larger argument is about reliability. Models that succeed only 45 percent of the time are unusable for tasks such as ordering food, booking airline tickets, sorting email, or filing invoices. Featherless AI says its action R1 agent beats Claude Sonnet, Gemini, and OpenAI on form-filling reliability, even though it is not aimed at PhD-level mathematics.
Upside turns scattered enterprise records into a usable data map
Jonas Bauer describes sales and marketing teams as data hoarders that store information in Salesforce without knowing what to do with it. Upside uses language models to extract useful details from messy records, such as sales emails, and put them into structured form. He compares the product with search engines and web crawlers that organize the unstructured web. The goal is a data command center built from a map of enterprise interactions, which teams can use to understand what reaches customers. Jonas says a public LinkedIn post produced many demo requests and customers waiting to access the platform.
OpenAudio makes voice generation controllable by delivery, not just words
Lengyue introduces OpenAudio after describing earlier work on Fish Audio and open-source voice models. The company's S1 model is presented as an instructible voice model. It lets users control how a phrase is spoken, including emphasis, focus, and yelling, rather than controlling only the text itself. The demonstration uses an emotional line to show those controls. Lengyue claims S1 is the most expressive voice model and says Fish Audio is available immediately at fish.audio.
OpenRouter reduces the cost of switching between language models
Alex Atallah describes OpenRouter as a marketplace and a single API for language models. Developers pay once and can move between models with little switching cost, while OpenRouter handles tool calling, caching, edge cases, and routing for price and performance. The product includes model comparison through a chat room, fine-grained privacy controls, usage observability, and public rankings based on real-world prompts. Alex says the company is trying to make a heterogeneous model ecosystem feel consistent, treating inference as a commodity. Its middleware, called plugins, can call MCPs and transform model outputs.
"We do all the heavy work to implement tool calling, edge cases, caching, and give you the best prices and performance possible for your region or wherever your servers are deployed."19:55
Who should watch
You are building AI products and want examples of startups focused on creative tools, voice interfaces, marketing, enterprise data, or model infrastructure.
Your agents work inconsistently on ordinary business tasks, and you want to hear the case for investing in reliability instead of larger models.
You are deciding whether to support multiple language models and want to see what a unified API, routing layer, and comparison product can provide.
