Pokee AI has just released Pokee-Isaac 28B, a 28-billion-parameter model that claims to be the world's first real 10-million-token context window — and it supposedly runs on a single graphics card you can actually buy. That is an enormous claim, so let me be straight with you from the first paragraph: every benchmark number below comes from Pokee's own internal testing, and nobody independent has verified it yet. Treat all of it as "Pokee says" for now. But if even half of it holds up, this is one of the most interesting releases of 2026.
📺 Watch: NEW Pokee-Isaac 28B Has a 10M Token Context!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
I have not run this model myself yet. This is my breakdown of what Pokee has published, why it matters, and exactly how I plan to put it to work the moment I get access.
What Is Pokee-Isaac 28B?
Pokee-Isaac 28B is an AI model from Pokee AI, released on 4 August 2026. At 28 billion parameters it is a medium-sized brain — not tiny, not huge, sitting in the same weight class as Qwen 3.8 27B, which I have covered before. On size alone, nothing unusual.
The headline is the context window: 10 million tokens, which works out at roughly 15 million words. That is a whole codebase, a full year of team chat history, or dozens of long contracts in one single prompt — no chunking, no retrieval tricks, no chopping your documents into pieces and hoping the model stitches them back together.
And it is built to be an agent, not just a chatbot. It plans, uses tools, and works through long tasks step by step — the same pattern behind agents like Prime. Give it a goal, hand it everything it needs to know, and let it work.
Why "Real" Context Is the Whole Story
Plenty of models advertise 1-million or 2-million-token contexts. But the number on the box and what the model can actually use are two very different things.
Pokee ran the RULER memory test: hide a fact deep inside a giant pile of text, then ask the model to find it. On Pokee's own runs, the other big models drop to a flat zero once you push past 2 million tokens. The window exists on paper; nothing is retrievable in practice. Pokee-Isaac, by their numbers, stayed above 93% all the way out to 10 million tokens, scoring 93.3 at the full 10 million.
That is the actual story here. Not a bigger window — a window that still works when it is full. If that survives independent testing, it quietly changes what "long context" even means.
If you want to put long-context models to work in your business instead of just reading about them, the AI Profit Boardroom walks through setups like this step by step inside the Agent OS. → Get the long-context playbook
It Runs on One Graphics Card
This is the part that made me lean forward. Pokee says Pokee-Isaac 28B runs locally on a single graphics card, starting from an RTX 4090 — a card a regular person can genuinely buy. One card. Not a cluster, not a data centre.
That opens up your own setup, a private cloud, office machines, even on-device use. It ships with day-zero support for the popular model-serving tools vLLM and SGLang, so if you have been through my Hermes local model setup guide, the deployment pattern will feel familiar.
And the privacy angle is enormous. Hospitals, law offices, government teams — anyone whose rule is "the data does not leave this building" — have been mostly locked out of this kind of long-context power until now. A model that reads ten million tokens without a single byte leaving your own machine is a genuinely new option for those people.
The Speed Number
A huge context is useless if it takes an hour to read. Pokee's figure: on a single B200 chip, the model reads context at up to 137,000 tokens per second — and, they say, it holds that even at the full 10 million. Their chip, their test, their number. But it is the right number to chase, because ingestion speed is what makes a 10-million-token prompt practical rather than theoretical.
How It Works, in Plain English
Most modern models are decoder-only designs. The more text you feed them, the faster their memory gets eaten — which is exactly why giant contexts get heavy, slow and eventually break down.
Pokee-Isaac uses what they call a non-decoder-only design: different plumbing that holds the context without collapsing and without demanding a giant machine. Pokee is keeping some of the details secret; the rest is in their technical report. I will not pretend to know more than they have published, but that architecture bet is the reason the 10-million claim is even on the table.
The Benchmarks: Where It Wins and Where It Loses
According to Pokee's own benchmark summary, the model:
- Leads on the memory tests — RULER and a harder one called MRCR.
- Leads on a function-calling test and on a multi-step agent test.
- Comes second on a terminal-task test, behind one of the big frontier models.
- Comes third on a tool-finding test.
Honestly? The losses make me trust it more. A model that wins some and loses some looks like a real set of results; a chart where the newcomer magically tops every single row is usually marketing. Second on terminal tasks against a frontier heavyweight is respectable for a 28B model.
There is a safety result too: on a red-team test — people actively trying to trick the model into doing bad things — it was the hardest of the six models tested to break, with the lowest attack success rate in the group.
The Big Caveat: These Are Pokee's Own Numbers
I keep repeating this because it matters. Every figure above comes from Pokee's own internal benchmark summary — it says so right on their chart — and no independent lab has verified the 10-million-token claim yet. Vendor benchmarks always show the product on its best day.
Which is exactly why, the moment access lands, this model goes straight into my Goldie Bench side-by-sides against the current frontier on real tasks. That is how self-reported numbers get checked: not by arguing about charts, but by running the same job on every model and watching what comes back. Until then my position is simple — take the claim seriously, hold it as "Pokee says".
📺 Watch: New PokeeClaw DESTROYS OpenClaw?
Cloud Giants vs Pokee-Isaac 28B
I have covered plenty of releases this year — GLM 5.5 among them — and the trade-off here is refreshingly clear. The huge cloud models can be smarter on some tasks. But most cannot 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.
| Question | Huge cloud models | Pokee-Isaac (local) |
|---|---|---|
| Raw intelligence | Can be smarter on some tasks | A little less brainpower |
| Usable long context | Big windows on paper; near zero past 2M on Pokee's RULER runs | 93.3 at the full 10M, per Pokee's own tests |
| Where your data goes | Someone else's servers | Stays on your machine |
| Hardware | None, but you rent access | From a single RTX 4090 |
| Best for | Peak reasoning on short tasks | Codebases, archives, privacy-locked work |
How I Would Put Pokee-Isaac 28B to Work
I have not got hands on it yet, so here is exactly what I will run first — three jobs straight out of my own business:
- Map the entire AI Profit Boardroom library in one prompt. Every lesson, every walkthrough, a full year of member questions — dropped in at once — then have it map a fresh content series from all of it together. Nothing I use today can see that whole library in one go.
- Audit the full member onboarding flow. Point it at the whole journey and ask it to find every spot where a new person might get stuck. One pass, complete picture.
- Mine the knowledge base for real questions. Feed it everything, pull out the questions members keep asking, then build tutorials around exactly those.
All three follow the principle I run across my Agent OS: give the agent everything, then let it plan. If the memory claims hold, this becomes the long-context engine in that stack — and a serious rival to the local models in my best open-source models for a Hermes agent rundown.
The Practical Caveats
- Self-reported numbers. Worth its own bullet even though I have said it three times already.
- Text-only for now. No images, no audio. Documents and code are the game.
- Not a free download. You do not grab it off a website. You use it through Pokee's API or deploy it yourself in your own setup, so plan for that.
- Vague in, vague out. A huge context still needs a clear ask. Ten million tokens of input plus a lazy one-line question equals a lazy answer.
Pokee-Isaac 28B: FAQ
Is the 10-million-token context actually real?
Per Pokee's own RULER runs, yes — above 93% at every depth, 93.3 at the full 10 million, while rivals hit zero past 2 million. But that is the vendor testing its own product. The honest answer today: credibly claimed, not yet proven.
Can I run it at home?
Pokee says yes — from a single RTX 4090 upwards, with day-zero vLLM and SGLang support. Consumer hardware, though not cheap consumer hardware.
Is it free?
No. It is not a download you grab off a website. You either use Pokee's API or deploy it yourself in your own environment.
What is it best at?
Long-memory work: big codebases, long documents, research archives, years of support tickets — and anything where the data cannot leave the building. Remember it is text-only.
When should I trust the numbers?
When people outside Pokee reproduce them. Until independent side-by-sides land — mine included — the 93.3 figure is a claim, not a fact. A promising claim, though.
My Verdict on Pokee-Isaac 28B
Most model launches are a bigger number on the same idea. This one is a different idea: a medium-sized model that bets everything on memory you can actually use, on hardware you actually own. The wins-and-losses benchmark chart reads honest, the red-team result is quietly impressive, and the local-first angle unlocks whole industries that were shut out entirely.
If the 10-million-token claim survives contact with independent testing, Pokee-Isaac 28B is not just another release — it is the first model where "give it everything" becomes a literal instruction. I will be testing exactly that the day I get access.
If you want to be ready to use models like this in your business the day they land, check out the AI Profit Boardroom — inside you get walkthroughs for running models on your own machine, live coaching calls where you share your screen and ask, ready-to-use long-context prompts, a 30-day roadmap, and 4,000+ members building together. → Join before your competitors do
