Meta Unveils an Open Version of Its Most Powerful A.I. Model

https://www.nytimes.com/2026/08/10/...4498&user_id=c383821527c441214d07ce6e4a6ba12a

Called Muse Glimmer, it is nearly identical to Muse Spark and can generate code, text and images. Muse Spark, which debuted in July as a “closed” A.I. model that people pay to access, will remain closed.

“Rather than centralizing superintelligence, we should distribute it,” Mr. Zuckerberg, 42, wrote in the essay, which he titled “The Future Is for Everyone.” He added, “This has the potential to begin a new era of personal empowerment.\

Muse Spark does not perform as well as leading A.I. models from Anthropic and OpenAI on measures such as coding, reasoning and writing, according to leaderboards that track A.I. benchmarks

Interesting.

I wonder how many parameters? Dense or MoE?

How much memory will one need in order to run it?

Edit:

Found answers myself. It is apparently a 30B class dense model, so it should be runnable on mere mortal hardware (albeit expensive mere mortal hardware)

My bet is that he's trying to get some good press right now. Between the recent lawsuits about suicides and the issue with his ship up in Alaska he's taking a beating in the press.

30b isn't big enough... "Most powerful model" without the knowledge.

I have lots to criticize Meta regarding in almost every way, but I've been supportive of their release of open AI models, ollama and the like. I have no doubt there's a vested interest one way or another, but it could end up being widely beneficial to the users and overall AI capability if allowed to be truly open and distributed, and the more they can pressure OpenAI / Anthropic and the like to do the same, the better. I do wish that Zucc and Meta would rather stop some of their other behavior if they really think openness is for everyone, then put their resources into legal battles over FOSS model training and IP law, instead of surreptitiously pushing every age verification data mining policy as a 'grassroots' facade.

mnewxcv said:

30b isn't big enough... "Most powerful model" without the knowledge.

Click to expand...

Well, I guess the question is how much of it has been distilled or otherwise acquired from larger models "institutional knowledge". We've already seen lots of progress in these regards with optimized versions of open larger models (ie DeepSeek) that can run on hardware much weaker than the full version yet have the majority of its capability.

mnewxcv said:

30b isn't big enough... "Most powerful model" without the knowledge.

Click to expand...

the big model is virtually free (not sure it cost more than the electricity to run it), if you do not mind your usage being used to help the next version too...

https://dev.meta.ai/docs/pricing-rate-limits/

That kind of pricing of near frontier full model will do way more to distribute and share capacity than anything local can I think.

Cached input is so cheap, that running a inferior 30b model on a mac studio will cost about the same in electricity alone if you have the average price rate and regular use case like agentic coding on a large codebase (both in pennies for a lot of work)... let alone the big upfront cost.

“Rather than centralizing superintelligence, we should distribute it,” Mr. Zuckerberg

Click to expand...

And how is he mining data from these models to benefit him... with all the crap against Meta right now for child abuse content, showing ad's to underage, directing underage kids to questionable content...

RanceJustice said:

Well, I guess the question is how much of it has been distilled or otherwise acquired from larger models "institutional knowledge". We've already seen lots of progress in these regards with optimized versions of open larger models (ie DeepSeek) that can run on hardware much weaker than the full version yet have the majority of its capability.

Click to expand...

TBF I didn't see a clear benefit in running larger versions of the same model, I think 25-30b is kind of a sweet spot it seems. Just because it is larger not necessarily means it will be more intelligent, pun intended. Sometimes more weights might just contribute more noise.

mnewxcv said:

30b isn't big enough... "Most powerful model" without the knowledge.

Click to expand...

Mostly agreed. That said, the current generation ~30b models are beating the previous generations 70B models on reasoning and accuracy through optimized weights somehow (I do not have the background knowledge to explain how they do it)

I'm just waiting for someone to release a ~65B-75B class model trained using these modern methods (preferably MoE so it runs faster) so I can take the full advantage of the 64GB on my dedicated AI GPU.

Right now everything I can seem to get my hands on is either modern 30B class or previous gen 70B class. The modern models are seemingly all 30B class or 120B class. There isn't much in between. The modern 30B class is beating the previous gen 70B class, but only by ~10%. Think of how great a modern 70B class model could be? I'm thinking something like Gemma4:26B MoE, but with 70B total parameters instead of 26b.

I can't help but wonder if the reason this hasn't happened yet is that they would be too good. They would likely be close enough to the frontier models that no one would have any reason to pay for the frontier models anymore. (either with money, or through spying+ad revenue)

RanceJustice said:

I have lots to criticize Meta regarding in almost every way, but I've been supportive of their release of open AI models, ollama and the like.

Click to expand...

They had their code stolen so they made their models open weight, because it was already out there. Don't give Pimpleberg more credit than he deserves. Lizard man.

They only want to open weight their models so others can compete. You see the same calls from other companies falling behind. They're not being altruistic.