tokennet
tokennet runs on machines people already own. If yours can hold a model in memory it can serve one, and earn credit for every token it produces.
Types of machines
A maker holds a whole model in memory and answers requests out of it. What decides which models you can serve is how much memory the weights can actually sit in — and machines provide that in two quite different ways.
Unified memory
One pool of memory that the CPU and the GPU both draw on. Most of it can hold model weights, so the number on the box is close to the number you can use. Plenty of room, moderate speed.
For example: a Mac mini, a MacBook Pro, a Mac Studio, an NVIDIA DGX Spark, an AMD Strix Halo mini-PC.
Discrete GPU
A graphics card with its own VRAM, separate from system memory. Only the VRAM counts: 64 GB of system RAM behind a 16 GB card still gives you 16 GB to put a model in. Less room, but that memory is typically several times faster.
For example: a PC with a GeForce RTX 5080 or 5090, a workstation with an RTX PRO 6000, a rig with several cards in it.
Both are first-class here. Unified memory is the one people overlook — "I don't have a GPU" is the commonest reason given for not joining, and it is usually said on a Mac that would serve perfectly well.
One exception, because it is on the chart below as a machine and not yet as a maker: AMD's Strix Halo is not supported. The relay knows how to budget a host like it, but the maker does not recognise that memory as unified and no recipe is sized for one, so such a machine would be offered far less than it can actually run. It is on the list.
How much memory do I need?
Not all of a machine's memory is available for weights. A Mac keeps a quarter of the pool back, a graphics card gives up about a gigabyte, and a DGX Spark holds back an eighth. What is left is what a model has to fit inside:
| Machine | Memory | Usable for weights | What it can serve |
|---|---|---|---|
| PC with one GeForce RTX 5080 | 16 GB | 15 GiB | gpt-oss-20b |
| Mac mini (M4) | 24 GB | 18 GiB | gpt-oss-20b, at its full 128K context |
| MacBook Pro (M4 Max) | 64 GB | 48 GiB | the 27B and 30B models, largest quantization, 256K context |
| Mac Studio (M3 Ultra) | 128 GB | 96 GiB | every model in the catalog with a Metal recipe |
| NVIDIA DGX Spark | 128 GB | 107 GiB | the 120B-class models |
The two 16 GB cases are worth reading together, because they are the one place the machine type changes the answer rather than the amount: 16 GB on a card leaves 15 GiB and fits a model, while 16 GB in a Mac leaves 12 GiB and fits nothing in the catalog. A Mac is the machine that needs 24 GB to start.
What memory bandwidth buys you
Memory decides which models you can run. Bandwidth decides how fast they answer. Producing a token means reading the model's weights out of memory — all of the active ones, once per token — so the ceiling on speed is roughly the machine's memory bandwidth divided by the size of the weights.
A 12 GB model on a machine with 120 GB/s of bandwidth tops out around ten tokens a second. The same model on a card with 960 GB/s is nearer eighty. Real figures land under the ceiling — attention and sampling cost something too — but the ratio is what separates two machines holding the same model.
That is the trade in the chart below. Unified memory buys capacity, a discrete card buys speed, and the machines that manage both are the expensive ones. All of them are useful to the network; they are useful for different models.