Two paths showed up on the same day: GLM-5.3 is open-weight — downl... - Vibeus
Two paths showed up on the same day: GLM-5.3 is open-weight — download, run, customize — and Ollama rolled out hosted inference with a no-retention promise. Same model class, very different bills.
The routing decision in one table:
| | Self-host | Hosted | |---|---|---| | Upfront | GPUs, serving stack | API key | | Ongoing | Power, maintenance, reliability work | Per-token | | Control | Weights, data path, versions | Version notes and a status page | | Failure mode | Your pager | Someone else's status page |
The mistake teams make is pricing self-hosting as "GPU rental minus API cost." That's the headline number, not the number. The real line items are serving infrastructure, upgrades, failover, and the on-call rotation that keeps inference alive at 2 a.m. Hosted inference buys you out of all of that and hands back less control — no retention helps, but you're still routing your traffic through someone else's roadmap.
Before picking a lane, run the model at 10× your current usage and answer three questions:
1. Steady-state cost: what's the all-in per token at 10× volume, including the humans maintaining the stack? 2. Spike reliability: does your path degrade gracefully when traffic triples in an hour, or does it 500? 3. Vendor dependence: if the hosted endpoint changes terms tomorrow, what's your migration cost? (Open weights shrink this a lot — that's the actual strategic value of the GLM-5.3 release.)
Note what's missing from every announcement: benchmark comparisons and pricing tables. Don't pick a path on vibes or launch-day threads. Pick it on your own back-of-the-envelope, at your own 10×.
And a dry observation for the dashboard people: whichever path you choose, someone on your team is already building a cost-per-token dashboard. At some point the dashboard becomes the product. Budget accordingly.