Open-weight vs closed AI models: what each one really costs

The open versus closed argument usually gets fought on principle. It’s more useful fought on what each one actually costs you, because the tradeoffs are concrete and they’ve moved this month.

Meta shipped Muse Glimmer, a 30-billion-parameter model under Apache 2.0, sized to run on consumer hardware. It also said weights for Muse Spark 1.2, its most advanced model, are coming.

That matters because the usual defence of closed models is a capability gap. If a lab open-sources its frontier model, that argument gets considerably harder to make.

	Open weights	Closed API
Cost shape	Fixed: hardware and ops	Variable: per token
Data residency	Yours entirely	Provider’s terms
Upgrades	You migrate, when you choose	Automatic, sometimes unannounced
Deprecation risk	None, you hold the weights	Real, models get retired
Ops burden	Yours	Theirs
Licence traps	Check: “open” is not always Apache 2.0	Terms of service

What open weight vs closed models actually cost

Start with cost, since it’s the one people get wrong in both directions. A closed API costs nothing until you use it, which is perfect at low volume and punishing in a loop.

Self-hosting inverts that. You pay for the GPU whether it’s busy or idle, so the question is simply whether your monthly token spend exceeds what the hardware costs to rent. Below that line, closed wins on economics alone.

Open weights are not free. They swap a bill you can see for an operations burden you will feel.

Nobody can take open weights away

The underrated advantage of open weights is that nobody can take them away. A closed model can be deprecated, repriced, rate-limited or quietly changed underneath you, and your evaluation results stop meaning anything the day that happens.

With weights on your own disk, the model you tested is the model you ship, indefinitely. For anything regulated or long-lived, that reproducibility is worth more than a few points of capability.

Licence terms are where “open” gets slippery, so read them. Apache 2.0, which Glimmer uses, is genuinely permissive. Plenty of models marketed as open carry usage caps, field-of-use restrictions or acceptable-use clauses that a closed API would never impose.

The capability argument, stated fairly

The closed side has a real answer on capability, and it should be stated fairly. The top of the Artificial Analysis index is still occupied by closed models, and a 30B model is not competing with them on hard reasoning.

Meta plans to open the weights on that most advanced model too, per Yahoo Finance, which is the commitment that would actually close the argument if it ships.

What has changed is the gap’s relevance. If your workload is extraction, classification, summarisation or routing, and most production workloads are, then a 30B model running locally is not obviously worse than a frontier model behind an API.

There’s also a policy dimension that closed advocates raise and it deserves acknowledging. Open weights cannot be recalled, and safety mitigations can be fine-tuned away by anyone with modest resources, which is the argument underneath current export control debates.

Zuckerberg’s counter, made in a 6,500-word essay alongside the release, is that concentrating powerful AI in a handful of companies carries its own risk. ABC News covered the argument, and CNBC read it as a swipe at OpenAI and Anthropic.

The full decision involves more than cost, and our comparison of hosted against self-hosted covers data residency, stability and the failure mode each side is prone to.

The practical answer

So the practical answer isn’t ideological. Start closed while volume is low and requirements are unclear, measure your actual token spend for a month, and revisit once the bill is real. TechCrunch framed Glimmer as a bet that the local option keeps getting harder to ignore, and on cost grounds that bet looks reasonable.