Open weights are back: Meta reopens the model debate, and buyers get leverage

Meta has swung back to releasing open-weight AI models, and it has not been quiet about it. Around the company's Q2 2026 earnings, Mark Zuckerberg took direct aim at OpenAI and Anthropic, arguing that a closed approach concentrates too much power in too few hands - and framing openness as the more American position, against what he characterised as a doom-focused mentality used to justify keeping the strongest models locked away.

The accompanying release is Muse Glimmer, an open-weight model that is deliberately much smaller than frontier systems, aimed at agentic tasks, and able to run on a Mac or PC with a single graphics card. Meta has also said it intends to release the weights of Muse Spark 1.2, the more capable model built by the superintelligence team it assembled last year.

Read the business model, not the manifesto

It is worth being clear-eyed about why this is happening. Meta does not sell model access. It monetises attention across its apps. OpenAI and Anthropic sell model access as the core business. Commoditising the model layer costs Meta very little and costs its rivals a great deal.

That does not make the argument wrong. It does mean buyers should treat "open" as a commercial strategy with real consequences for them, rather than as a philosophical stance to agree or disagree with.

The question that actually matters in procurement

For most organisations the choice is not open versus closed in the abstract. It is: does controlling the weights change anything for this workload?

It genuinely does when:

• Data cannot leave. Regulated, contractual or sovereignty constraints sometimes make an API a non-starter regardless of the vendor's assurances.
• You need version stability. Hosted models change underneath you. If you have validated a workflow against specific behaviour, a model that cannot be silently updated is worth a lot.
• The workload is narrow and high-volume. A small model doing one job well, close to your data, can beat a frontier API on both latency and unit cost.
• You need to survive a supplier. Weights you hold are the only real continuity plan if a vendor changes pricing, terms or direction.

It usually does not when the workload is varied and reasoning-heavy. Frontier hosted models still lead there, and the gap is expensive to close yourself.

The economics are less flattering than the pitch

The part that tends to get skipped is total cost. Open weights are free; serving them is not. Once GPU capacity, redundancy, evaluation, monitoring and the engineers who own all of it are counted, the break-even against per-token API pricing is frequently much further out than expected - and for low-to-moderate volumes it may never arrive at all.

That is not an argument against open weights. It is an argument for doing the arithmetic against your actual usage before treating self-hosting as the cheaper option, because "free model" and "cheaper system" are different claims.

A more useful way to compare

On TrustList's AI model rankings we score models on a blend of published benchmark capability, review sentiment and community signal. Benchmarks tell you what a model can do in ideal conditions; they do not tell you what it costs to run, how stable it is, or whether the licence permits your use case.

So when comparing an open-weight option against a hosted one, add the questions benchmarks never answer:

• What does the licence actually permit commercially? "Open weights" is not the same as "open source", and several popular licences carry real restrictions.
• Who patches it, and how quickly, when a problem is found?
• What is the serving cost at your volume, including the staff time?
• If you stopped paying your current vendor tomorrow, what would break?

The genuine win

The most valuable outcome of Meta's move is not any single model. It is that a credible open tier at every capability level gives buyers a negotiating position they did not have when the only choice was which closed API to depend on.

Most organisations will still buy hosted models for most work. They will simply do it on better terms - and that is worth more to buyers than either side's rhetoric.

---

Sources: BNN Bloomberg on Meta's open-weight push and Muse Glimmer · Crypto Briefing on Zuckerberg's criticism of closed rivals