Pokee-Isaac 28B claims the world’s first genuinely usable 10 million token context — and it runs on a single RTX 4090. Here’s what the benchmarks show, where it loses, and the caveat you need before you plan anything around it.
Short answer
- 28B parameters from Pokee AI, released 4 August 2026, built as an agent not a chatbot.
- On Pokee’s RULER tests, rivals hit zero past 2M tokens; Pokee-Isaac held 93.3% at 10M.
- Runs on a single graphics card from an RTX 4090, with vLLM and SGLang support.
- All benchmarks are Pokee’s own, and it’s not a free download — API or self-deploy.
What Pokee-Isaac 28B is
Pokee-Isaac 28B is an AI model from a company called Pokee AI. They released it on 4 August 2026. The 28B means 28 billion parameters — a medium-sized brain. Not tiny, not huge.
What makes people stop scrolling is the claim attached to it: Pokee call it the world’s first real 10 million token context model.
10 million tokens is roughly 15 million words. A whole codebase. A full year of your team’s chat history. Dozens of long contracts, all in one prompt. No chopping into pieces, no clever search tricks to feed it bit by bit. You hand it the whole thing.
It’s also built as an agent, so it doesn’t just answer questions — it can plan, use tools and work through long tasks step by step.
Why they keep saying the word “real”
Here’s the uncomfortable secret about context windows. Plenty of models advertise a million or two million tokens. The number on the box and what the model can actually use are two different things.
Pokee ran a memory test called RULER — think of it as hiding a fact deep inside a giant pile of text and asking the model to find it.
| Context length | Other major models | Pokee-Isaac 28B |
|---|---|---|
| Under 2M tokens | Works | Works |
| Past 2M tokens | Drops to a flat zero | Stays above 93% |
| Full 10M tokens | Not usable | 93.3% |
Zero. As in the window exists on paper, but the model can’t pull anything back out of it.
That’s the whole story. It isn’t just a bigger window — it’s a window that still works when it’s full.
🔥 Want this set up without the guesswork? If you want to know which card you need, how to feed it that much text without it choking, and how to keep it private on your own machine — that’s exactly what we walk through together. Inside the AI Profit Boardroom you get the Agent OS as a ready-to-install file, a 30-day roadmap, daily tutorials the same day new tools ship, and four live coaching calls a week where you share your screen and get unstuck. → Get access here
It runs on a single graphics card
This is quietly the bigger deal. Normally a model holding this much memory needs a room full of heavy machines wired together, which is why this kind of power has been stuck in the cloud — you send your data to someone else’s servers and hope for the best.
- Pokee say it runs on a single graphics card starting from an RTX 4090. One card, not a cluster.
- You can run it in your own setup — private cloud, office machines, even on-device.
- Day-zero support for vLLM and SGLang, the tools people already use to serve models.
- On a B200 it reads through context at up to 137,000 tokens per second, even at the full 10 million.
For anyone working with client files, member data or anything sensitive, that last point about the data never leaving the building is the whole pitch.
How it pulls this off
Pokee are keeping some of it secret, but they do share the key idea.
Most models today use a decoder-only design. It works well, but it has a nasty habit: the more text you give it, the more memory it eats, and fast. That’s the real reason 10 million tokens usually gets so heavy that everything breaks down.
Pokee-Isaac uses what they call a non-decoder-only design. Different plumbing underneath. That’s the trick that lets it hold all that context without falling apart or needing a giant machine.
Where it wins and where it doesn’t
| Test | Result |
|---|---|
| RULER (memory) | Leads — not close |
| MRCR (harder memory) | Leads |
| BFCL v4 (function calling) | Leads |
| Tau-cubed Bench (multi-step agent) | Leads |
| Terminal task test | Second, behind a big model |
| Tool finding | Third |
| DTAP (red-team safety) | Hardest of six models to break — lowest attack rate |
So it’s strong, but it doesn’t sweep everything. Honestly, that mix makes me trust the numbers more, not less. A model that wins some and loses some looks a lot more real than one that magically tops every row.
The caveat that matters most. These numbers all come from Pokee’s own testing — it says so on their chart, “internal benchmark summary”. That doesn’t mean they’re fake. It means nobody outside has checked them yet. Take the long-context claim seriously, but treat it as Pokee’s claim until someone independent runs the same tests.
Who it’s actually for
If you deal with huge piles of text, this is for you.
- People working with big codebases.
- Anyone buried in long documents — contracts, research, reports.
- Support teams sitting on years of tickets.
- Anyone tired of copying and pasting in chunks and watching the AI forget the beginning by the time it reaches the end.
And it’s especially for anyone who can’t send data to the cloud — hospitals, law offices, government teams, anywhere the rule is that the data does not leave this building. Until now they were mostly locked out of this kind of power. A model that runs on one card in your own office changes that.
Three things to know before you try it
- It’s text only right now. No images, no audio. Don’t hand it a picture and expect magic.
- It’s not a free download. You use it through their API or deploy it yourself in your own setup. Plan for that.
- Be specific with giant prompts. A huge context is powerful, but a clear ask still wins. Tell it the exact thing to find or build — vague in, vague out.
Quick comparison so you know where it sits: the huge cloud models can be smarter on some tasks, sure. But most of them can’t actually use their big windows once full, and you have to trust someone else with your data. Pokee-Isaac trades a little raw brainpower for a memory that really works, on a machine you control.
Want help running models like this? The AI Profit Boardroom has walkthroughs for getting models running on your own machine step by step, ready-to-use prompts for long-context work, four coaching calls a week and a roadmap so you’re building a workflow rather than collecting tools. Start free with the free AI course and community.
FAQ
What is Pokee-Isaac 28B?
A 28-billion-parameter agentic AI model from Pokee AI, released 4 August 2026, claiming the first genuinely usable 10 million token context window.
Is the 10 million token context real?
On Pokee’s own RULER testing it held above 93% out to 10M, scoring 93.3 at full length, while other major models dropped to zero past 2M. Those numbers are self-reported and not independently verified yet.
What hardware do I need?
Pokee say it runs on a single graphics card starting from an RTX 4090 — one card, not a cluster — with day-zero support for vLLM and SGLang.
Can I download it for free?
No. It’s not a free download — you use it through their API or deploy it yourself. Budget for that before you plan around it.
Does it handle images or audio?
Not right now. It’s text only, so feed it text.
How is it different architecturally?
Most models use a decoder-only design whose memory use balloons with context length. Pokee-Isaac uses a non-decoder-only design, which is what lets it hold 10M tokens without needing a cluster.
Is it good at agent tasks?
It leads on function calling and multi-step agent benchmarks in Pokee’s tests, but comes second on terminal tasks and third on tool finding. Strong, not a clean sweep.
Who benefits most from it?
Anyone with huge text piles — big codebases, long documents, years of support tickets — and especially regulated industries where data cannot leave the building.
The bottom line
Pokee-Isaac 28B matters less for its size than for proving a point: a context window is only worth what you can actually pull back out of it. If the independent tests hold up, a model that keeps 93% recall at 10 million tokens on a single RTX 4090 changes what’s possible for anyone whose data can’t go to the cloud. Treat the numbers as Pokee’s until someone else runs them.
About Julian Goldie
I run Goldie Agency, a 7-figure SEO agency, and teach this daily on a 400K+ subscriber YouTube channel. 240+ client projects on Upwork at a 100% job-success score, 10+ years through every major Google update. My systems are in the AI Profit Boardroom; my link building book is free here.
