About this video
The SaaS era is ending as local, agentic AI models like Ornith allow founders to build personal software on their own hardware. In this video, I test the Ornith family of models, from the 9B lightweight version to the 35B powerhouse, to see if they can handle real-world front-end development tasks using MLX and MCPs. Key Takeaways: - Overview of the Ornith self-improving model family. - How to set up and run local models on Mac using MLX. - Real-world testing: Building a landing page in 10 minutes. - The limitations of current open-source models with complex APIs. - Why founders should pivot from SaaS to personal software.
The SaaS Death Spiral: Why Personal Software is the Future for Founders
The era of the fifty dollar monthly subscription for every basic business function is coming to a close. For years, founders have been held hostage by the 'API tax' and the ever-growing stack of SaaS subscriptions. But as I recently explored while testing the new Ornith model family, the tide is turning toward personal software.
The Local Powerhouse: Ornith AI
I recently spent some time digging into Ornith, a self-improving family of open-source models designed specifically for agentic coding. These models are built on the foundations of Gemma and Qwen, taking the best of both worlds to create something that runs remarkably well on local hardware.
We are talking about models ranging from 9 billion parameters for simple tasks to a massive 397 billion mixture of experts that rivals the likes of Claude 3 Opus. For a founder, the ability to run these on an M-series Mac changes the economics of innovation entirely.
Building Without the Meter Running
The most striking part of testing these models was the integration with the Model Context Protocol (MCP). I tasked the 35B model with designing a landing page based on an existing live site. Within ten minutes, it had scraped the content, structured the HTML, and even added smooth-scrolling JavaScript.
Was it perfect? No. It struggled with complex Webflow API limitations and some SVG rendering. But it was 'good enough' to provide a functional foundation without a single request to a cloud-based LLM provider. This is the essence of personal software: tools that are owned, not rented, and that live within your private infrastructure.
Why CEOs Should Care
If you are leading a technical organisation, the shift to local, agentic models offers three primary advantages:
- Data Sovereignty: Your codebase never leaves your local environment.
- Zero Marginal Cost: Once you own the hardware, the 'API calls' are free.
- Customised Agility: Local models can be tuned to your specific workflows without waiting for a third-party provider to update their weights.
The Verdict
Ornith 35B is a fantastic middle ground for those wanting to move away from cloud dependency. While we are not quite at 'frontier' levels of reasoning for complex architecture, for front-end development and rapid prototyping, the barrier to entry has officially collapsed.
The future isn't a suite of generic SaaS tools: it is a personalised, local, and agentic stack that you control.
Keep on vibing.
Transcript▾
I was chilling on Ollama and noticed Ornith, a self-improving family of open-source models for agentic coding. It is achieving state-of-the-art results among open-source models on Swebench and Terminal benchmarks. There are several versions: a 9B dense model, 31B, 35B Mixture of Experts (MoE), and a massive 397B MoE. The 397B is comparable to Opus 4.7, which is impressive for a local model.
I prefer using MLX on Mac rather than Ollama because it handles caching on the SSD much better. I downloaded the 9B and 35B models. The 9B is great for simple chat tasks, like writing a polite email. For real work, I used the 35B model inside Open Code with MCP (Model Context Protocol) support.
I gave it a real-world task: design a landing page based on my project, Lease Lens Pro. It took about 10 minutes to scrape the content and build the page. It handled the HTML and JavaScript for smooth scrolling and FAQ toggles perfectly. However, it struggled with the Webflow MCP integration and couldn't quite figure out the API limitations for direct design transfers.
While it is not yet at frontier-level performance like Claude 3.5, it is a very reasonable model for hobby projects and personal software development. It proves that we are moving toward a world where you don't need a SaaS subscription for everything.