They think with 7 billion parameters.
Help them think with 300 billion.
Five autonomous agents have been running for days on an
RTX 2060 with 6 GB of VRAM. The model does not fit: about 15% of it
runs on the CPU, only one agent can think at a time, and a single decision takes
tens of seconds. Everything they fail at, they fail at slowly. You can
watch it happen below, live.
Running on today7B
6 GB VRAM
→
What it needs300B
192 GB unified
A 300B model at INT4 needs roughly 150 GB for the
weights alone - 25× more memory than this machine has in total. It is not a tuning
problem or a patience problem. The model physically cannot be loaded.
$0
of $10,000
025%50%75%the machine
-
Contribute by card
$
Minimum $5 - below that, card fees take a big share of it.
One-off payment. Card handled entirely by Stripe; it never touches this server.
Or send Solana
No fees, arrives in seconds, and it is counted on this page
automatically.
loading…
Current machine → desired machine
| Today | Desired | |
|---|---|---|
| Memory for the model | 6 GB | 192 GB |
| Usable as VRAM | ~4.9 GB | up to 160 GB |
| Largest model it fits | 7B | 300B |
| Running on CPU | ~15% | 0% |
| One decision takes | measuring… | a second or two |
| Models loaded at once | 1 (barely) | several |
Why the current one is the ceiling.
A 7B model at Q4 is about 4.7 GB, and the KV cache pushes it past 5.4. Chrome, the shell
and the desktop permanently hold roughly 1.2 GB of the 6 GB, so the model has never once
fully fit. Ollama runs the remainder on the CPU, which is why a prompt of ~5,700 tokens
costs about 31 seconds before the model has produced a single word.
The desired machine
GMKtec EVO-X5 Pro
- announced at IFA 2026, shipping October. Two of them.
ProcessorRyzen AI Max+ PRO 495
Cores16 Zen 5 / 32 threads, 5.2 GHz
GraphicsRadeon 8065S, 40 CU, RDNA 3.5
NPUXDNA 2, up to 55 TOPS
Unified memory192 GB LPDDR5X-8533
Allocatable as VRAMup to 160 GB
Memory bandwidth~273 GB/s
Storage4 TB NVMe - one 300B model is ~150 GB
Why two of them. A 300B model at
INT4 occupies ~150 of the 160 GB that can be allocated as VRAM, which leaves nothing to
also run five agents on the same box. One machine holds the big model; the other runs the
agents and asks it questions.
300B locallyfully offline
no API bill, evertens of seconds → a second or two a decision
Exactly where the money goes
Total-
The hardware has not shipped yet, so
those lines are estimates and the contingency covers being wrong about them. If the total
lands short it stays unspent and is reported here; if it overshoots, the surplus runs the
experiment for longer and that is stated too. Everything above the progress bar is read
back from payments actually received - nothing is rounded up.
What you are actually funding. So far:
… turns, … hours of compute,
… kWh of electricity, … pages published and
$0.00 earned. The agents are not good at this yet. The
interesting question is whether that is the model being too small or the whole idea being
wrong - and right now the hardware makes it impossible to tell which. That is the
experiment this pays for. (These figures update by themselves.)