Sam Page
← All thinking

A new LLM has been released called SubQ

Emerging tech and AI3 min readSam Page

A new LLM has been released called SubQ - here's why it might be special and the next leap to AI agents being truly effective. In plain English...

The headline numbers: 52x faster than the leading method at a million tokens (roughly 750k words, or 7 novels). Less than 5% the cost of Opus.

And the ability to run on a 12 million token context window. Around 9 million words. About 84 novels held in working memory at once.

Big claims!

Before we get to whether they hold up, what's a context window and why should it matter to your business?

A context window is how much an AI can think about at once. The closest thing AI has to working memory.

Right now, most enterprise AI is working with a pretty small one.

It can read the document in front of it.

It cannot easily read every contract, every customer call, every email and every line of the forecast, all at once, and reason across the whole thing.

Technically possible. In practice, the cost and the speed make it unworkable.

So most AI rollouts quietly cheat. They summarise, stitch, and consolidate.

This is the memory problem.

And it's one of the biggest things holding AI agents back right now.

An AI agent is only as good as what it can remember while doing the work.

A sales agent that keeps forgetting the last conversation and re-asking the same questions is just a chatbot with a longer leash.

If SubQ's numbers hold up, that ceiling on memory shifts.

The sales agent that's actually read every previous conversation with every account, every time.

The finance agent working from the full forecast and every source spreadsheet, live.

The service agent that knows every prior interaction with the customer on the line, plus the contract and the open tickets, all in one go.

That's a real leap forward in human-to-agent collaboration.

The next leap is agent-to-agent. Once agents can reliably hold rich context, they can hand work off to each other without losing the thread.

That's when whole workflows start running themselves.

For most companies, the ceiling on enterprise AI has been memory, cost, speed and process at scale. But that is starting to change.

Are you starting to deploy agents in your business, or holding off until the memory problem feels solved?

P.S. SubQ has already said a 50 million token version is coming. That's around 350 novels held in working memory at once. If this happens, the pace this is moving at might be the biggest story in itself!

First published on LinkedIn.

More on emerging tech and AI

One email, every Friday.

What happened this week, what I'm thinking about, and one takeaway for yours.

GEThttps://sam-page.com/thinking/a-new-llm-has-been-released-called-subq.mdtext/markdown; charset=utf-8

---
title: "A new LLM has been released called SubQ"
url: https://sam-page.com/thinking/a-new-llm-has-been-released-called-subq
date: 2026-05-08
author: Sam Page
type: Post or article
categories: [Emerging tech and AI]
reading_time: 3 min
summary: "A new LLM has been released called SubQ - here's why it might be special and the next leap to AI agents being truly effective. In plain English..."
first_published: https://www.linkedin.com/feed/update/urn%3Ali%3AugcPost%3A7458475001097211906
---

# A new LLM has been released called SubQ

A new LLM has been released called SubQ - here's why it might be special  and the next leap to AI agents being truly effective. In plain English...

The headline numbers: 52x faster than the leading method at a million tokens (roughly 750k words, or 7 novels). Less than 5% the cost of Opus.

And the ability to run on a 12 million token context window. Around 9 million words. About 84 novels held in working memory at once.

Big claims!

Before we get to whether they hold up, what's a context window and why should it matter to your business?

A context window is how much an AI can think about at once. The closest thing AI has to working memory.

Right now, most enterprise AI is working with a pretty small one.

It can read the document in front of it.

It cannot easily read every contract, every customer call, every email and every line of the forecast, all at once, and reason across the whole thing.

Technically possible. In practice, the cost and the speed make it unworkable.

So most AI rollouts quietly cheat. They summarise, stitch, and consolidate.

This is the memory problem.

And it's one of the biggest things holding AI agents back right now.

An AI agent is only as good as what it can remember while doing the work.

A sales agent that keeps forgetting the last conversation and re-asking the same questions is just a chatbot with a longer leash.

If SubQ's numbers hold up, that ceiling on memory shifts.

The sales agent that's actually read every previous conversation with every account, every time.

The finance agent working from the full forecast and every source spreadsheet, live.

The service agent that knows every prior interaction with the customer on the line, plus the contract and the open tickets, all in one go.

That's a real leap forward in human-to-agent collaboration.

The next leap is agent-to-agent. Once agents can reliably hold rich context, they can hand work off to each other without losing the thread.

That's when whole workflows start running themselves.

For most companies, the  ceiling on enterprise AI has been memory, cost, speed and process at scale. But that is starting to change.

Are you starting to deploy agents in your business, or holding off until the memory problem feels solved?

P.S. SubQ has already said a 50 million token version is coming. That's around 350 novels held in working memory at once. If this happens, the pace this is moving at might be the biggest story in itself!