← All articles

Tokens, time, and knowing when local isn't the right fit

Not every job needs a whole box for an hour, and not every job belongs on a local model. This post covers the two ways to use Hypermesh today—per-token access and time-based leases—and our plan for the cases where a local model isn't the right fit.

Two ways to pay for compute

Per-token access (Full Model). You send requests to a model from our catalog that is running on an enrolled host, and you pay for the tokens you use. This works well for bursty or light workloads where you don't want to keep hardware warm. The platform take on per-token access is 18%.

Time-based leases. You rent a hardware class for a block of time and run what you need on it. This fits steady workloads, long batch jobs, and anything where you care about having the box to yourself. Hourly leases carry a 12% take, and so will Your Model once it ships. Your Model lets you bring a model by URL or upload, and it's coming soon.

A rough rule: if you would keep a leased box busy most of the hour, a lease usually costs less. If your traffic comes in short bursts, tokens usually cost less. Our post on tokens per second per user explains how sharing affects throughput, and true pricing explains where the floor under both prices comes from.

When a local model isn't a fit

Open models running on local hardware cover a lot of everyday work. Some requests still need more than that: very long contexts, the newest frontier reasoning, specialized modalities, or a model that is too large for any host class online at the moment. Saying no to those requests, or quietly serving a worse answer, isn't a good outcome.

What we're planning: smart routing

We plan to add automatic model selection and routing. When a request isn't a good fit for the local catalog, it could be sent to frontier and partner providers instead, with the routing decision visible to you. Routing itself is not live yet. Today you pick the model, and Full Model requests run on enrolled hosts.

Our goals for routing:

  • Local first, when it fits. Hardware that already exists and is otherwise idle gets the work it can do well.
  • Visible decisions. You should be able to see where a request ran and why.
  • Opt-in controls. If you don't want requests to leave enrolled hosts, you'll be able to say so.
  • Honest pricing. Routed requests will show their own posted price. We won't hide a partner's cost inside a local rate.

We'll publish details, including pricing for routed requests, before routing goes live.

Try it

Hypermesh connects hosts who have enrolled, checked hardware with renters who need time or tokens on it. Renters pay in USDC (or buy USDC through Stripe’s onramp), and hosts are paid through Stripe Connect. Take rates are 12% on hourly leases and Your Model, and 18% on per-token Full Model access. Hosts get weekly bank payouts once they reach the $5 minimum. Start at hyperme.sh.


Smart routing and Your Model are planned features; their details may change before launch. Per-token Full Model access and time-based leases are live product.