Models of your own, at research-team quality.

Pocket Model lets any team build its own AI models: an AI agent does the research and post-training, a person directs it, and the team ends up with a model it owns and can run anywhere.

Why I'm building this

I've been building with open models since Meta released OPT, and then LLaMA set off the explosion of open-weight AI and the communities around it. Alongside that, I've seen and been part of the labs where the frontier gets built.

The frontier plays in a different goal space. The major labs' focus is making their models safe, reliable drivers of long-horizon agents: agents that write software, operate computers and do deep knowledge work.^1,2

A few years ago the value of an LLM looked different: short context, short bursts, high throughput. LLMs quietly replaced a lot of traditional machine learning, like classifying a ticket, pulling fields from an invoice or reading a scanned page. Those jobs are short, repetitive, and run millions of times.

As the labs moved toward agents, fine-tuning as a service moved out of their focus. Across the major labs' APIs it has been removed, retired, deprecated or closed to new work.^3,4,5,6

Meanwhile, open-source and open-weight models, base and tuned, from labs in the US and in China, have become good enough to tune for exactly these jobs, and cheap enough to run yourself.⁷

Running it yourself matters more than it used to. Customers weigh more than a general benchmark of agentic performance: where their data lives and who can see it, what the infrastructure costs, and what their regulators require. Personal data carries its protections across borders,⁸ health data needs written guarantees from anyone who handles it,⁹ and the EU AI Act now applies.¹⁰ In one 2026 survey, nearly 60% of AI leaders named cross-border data restrictions as a major challenge.¹¹ Those guarantees can be hard to get from a shared, hosted inference service, and are straightforward with a model you run yourself.

Not everything is a nail, and we shouldn't automatically reach for the biggest hammer.

Over the next couple of years I expect a renaissance on two fronts: small language models, each built for one job, and traditional ML, now that frontier agents can do the research and engineering work that used to take an ML team.

Pocket Model sits where those meet. Your agent does the research and the training, you direct it, and you end up with a small model of your own that does your job better, faster and cheaper.

— Aleksey

Who's behind it

Aleksey Smolenchuk, founder
Second-time founder. Co-founded Crixet, acquired by OpenAI; 17+ years building production systems at Uber, Netflix and OpenAI. Has worked on post-training open-weight language models since the first releases, and Pocket Model is the product of that work.

lxe.co @lxe @lxe

Get in touch

Pocket Model, Inc. Questions, a task you'd like to try, or just hello: info@pocketmodel.ai.

Why I'm building this

Who's behind it

Get in touch

Sources