Quick answer. Meta reversed course on open weights in August 2026. It shipped Muse Glimmer, a 30B Apache-2.0 model, on August 10, and Mark Zuckerberg announced Muse Spark 1.2's weights will be open-sourced too — five days after launching the closed, paid Muse Code agent. The Spark 1.2 weights have not shipped yet.
In April 2026, Meta ended the Llama era and replaced it with a closed model family. On August 5 it doubled down, launching Muse Code — a paid, closed coding agent with a tier that charges you less if Meta can train on your source code.
Five days later it released a fully open model and announced it would open-source the flagship too. This is one of the sharper strategy reversals in recent AI, and it is worth understanding rather than just noting.
What actually happened?
Three things, in eight days.
| Date | Event | Open or closed |
|---|---|---|
| Aug 5, 2026 | Muse Code + Muse Spark 1.2 launch | Closed, paid |
| Aug 10, 2026 | Muse Glimmer released, 30B Apache 2.0 | Open weights |
| Aug 10, 2026 | Zuckerberg announces Muse Spark 1.2 weights will open | Announced, not shipped |
The announcement came with a roughly 6,500-word essay arguing for American open-source AI, and it was picked up by CNBC, the Financial Times and Fortune. Glimmer itself went to the top of Hacker News with 1,199 points.
One thing to be precise about: the Muse Spark 1.2 weights are not released. As of mid-August 2026 Meta's Hugging Face organisation holds only the Glimmer repositories, and no date has been given. Anything describing Spark 1.2 as open right now is wrong.
Why would Meta give away a frontier model?
Four reasons, and they compound.
The closed strategy was not winning
Meta's own launch benchmarks put Muse Spark 1.2 second to Claude Opus 5 on all three coding evaluations it selected and ran itself. A fourth chart placed it fourth of six. When your closed flagship cannot lead the benchmarks you chose, charging premium prices for exclusivity is a weak position. We covered the numbers in detail in our benchmark analysis.
Meta has no enterprise-software business
Anthropic and OpenAI have spent years building developer sales motions, enterprise agreements and trust. Meta has none of that. Competing on the same closed-API terms means fighting on the incumbents' turf with no distribution advantage. Open weights are a distribution strategy for a company that owns no distribution channel to developers.
The economics are brutal
Meta's Q2 2026 results, reported on July 29, showed free cash flow collapsing to $784 million from $8.55 billion a year earlier, with capital expenditure of $31.1 billion and full-year guidance raised to $130–145 billion. Zuckerberg's stated rationale — that there is significantly higher margin in selling intelligence than selling compute — cuts both ways. If you cannot win the premium tier, commoditising it damages competitors more than it damages you.
Open weights buy back the ecosystem
Llama's real asset was never benchmark leadership. It was that a generation of developers built on Meta's models by default. The closed pivot forfeited that, and the community reaction was hostile. Glimmer's uptake — roughly 750,000 downloads across repositories within days — suggests the goodwill is recoverable.
Is Muse Glimmer actually competitive?
In its size class, partly — and Meta was refreshingly honest about where.
Against Qwen 3.6 27B, its closest open competitor, Meta's own published table shows Qwen winning four of seven benchmarks, including SWE-Bench Verified (77.2 vs 76.0), TerminalBench 2.1 (60.7 vs 51.7) and OSWorld-Verified (75.6 vs 65.9).
Glimmer's wins cluster around tool use: MCP Atlas (75.5 vs 62.5), WildClawBench (47.6 vs 43.2) and SWE-Bench Pro (51.2 vs 50.2). Against Gemma 4 31B it wins every row.
So Glimmer is a genuine tool-calling specialist rather than a general leader. Full details in our Muse Glimmer guide.
The strategically important point is not Glimmer's ranking. It is that a well-resourced American lab is once again shipping competitive open weights under Apache 2.0, after eighteen months in which the best open models came predominantly from Chinese labs.
What does this mean if you build on these models?
Do not rewrite your stack around an announcement. Spark 1.2's weights are not out. Plan against what has shipped.
Glimmer is worth evaluating now if you build agents. Apache 2.0, 128K context, fits a 24GB GPU at 4-bit, best-in-class MCP tool calling for its size. Low-risk to test.
Watch the price floor. Muse Spark 1.2 at $1.25/$4.25 per million tokens already undercuts Claude Opus 5 by roughly 5.9x on output. If those weights do open, the floor for self-hosted frontier-adjacent capability drops again.
Treat the closed products' data terms separately. Meta going open on weights does not change Muse Code's contributor tier, which still trades a 12–21x discount for training rights over your prompts and completions. See what that actually costs you.
The honest read
This is not a philosophical conversion. It is a company that tried the closed strategy for four months, found it could not lead on capability, and returned to the approach where it had structural advantages — scale, compute and a willingness to give away what competitors sell.
That is fine. Strategy driven by competitive position is more predictable than strategy driven by principle, and the second-order effect is real: whatever the motive, there are now more good open models than there were in July, and one of them is Apache 2.0 from a lab with Meta's resources.
The question worth watching is whether the Spark 1.2 weights actually ship, and on what license. Until they do, this is a promising announcement attached to one real, medium-sized release.
FAQ
Did Meta open-source Muse Spark 1.2?
Not yet. Zuckerberg announced on August 10, 2026 that the weights will be open-sourced, but they have not been released and no date has been given. Only Muse Glimmer's weights are actually available.
What is Muse Glimmer?
Meta's 30B open-weights multimodal model, released August 10, 2026 under Apache 2.0. It has a 128K context and runs in under 20GB at 4-bit quantization.
Why did Meta go back to open weights?
Its closed flagship could not lead the benchmarks Meta itself selected, it has no enterprise-software sales motion to compete with Anthropic and OpenAI, its capital expenditure is enormous, and open weights rebuild the developer ecosystem Llama once had.
Is Muse Glimmer better than Llama 4?
They target different niches — Glimmer is a 30B agentic tool-use specialist, while Llama 4 was a broader family. Glimmer is the current, actively supported open release; Llama is no longer Meta's forward line.
Does this change Muse Code's pricing or data terms?
No. Muse Code remains a closed paid product, and the contributor tier still exchanges a 12–21x discount for permission to train on your prompts and completions.
Should I wait for Muse Spark 1.2's weights before choosing a model?
No. There is no release date, and plans announced without dates frequently slip or change licence terms. Choose from what has shipped.