Meta's Muse Glimmer — Full Breakdown & Transcript

[00:01] last year starving Llama, poaching researchers with nine figure Zuck bucks, and locking its new models behind a Facebook auth page, just released a free and open source model under the Apache 2.0 license. It's called Muse Glimmer, a

[00:14] name I believe was inspired by Zuck's fan fiction My Little Pony character, and it's a 30 billion parameter agentic model that's small enough to run on your poverty spec PC. That means you can now have an always-on mini Zuck agent living

[00:27] rent-free in your gaming PC, reading your email, managing your calendar, and elections in Middle America. And according to Meta, an agent like that which is a sentence they've been dreaming about writing since 2004.

[00:42] Except this time, the surveillance can run entirely on your own hardware. In today's video, we'll find out what Muse Glimmer Glimmer can actually do, learn distillation, quantization, and speculative decoding, and determine

[00:55] whether Zuck's open-source redemption arc is sincere or if this is just what losing looks like right now. It is August 12, 2026, and you're watching The Code Report. It wasn't long ago when Meta was seen as the champion of the

[01:07] open weight class. Their Llama models slopped up over a billion downloads and created an entire ecosystem of custom models built on top of them. That is until Llama 4 was released, and it became the biggest Facebook bust since

[01:19] underage girls. In the release were two models, Scout and Maverick. Initially, because Meta submitted a secret juiced-up version to Llama Arena that

[01:31] download. And the models we could download had the intelligence of farm animals, which was fitting given their names. And while this was happening, Comrade Xi Jinping was dropping open models that were better, cheaper, and

[01:43] released in accordance with Mao Zedong's thought. Then Zuck responded to this humiliation the only way a billionaire knows how. He spent 14 billion on a 49% stake in Scale AI just to acquire its CEO, Alexander Wang. Then he went on a

[01:57] poaching spree for researchers at OpenAI and Google and rebranded the whole operation as Meta Superintelligence Labs. The first thing his new dream team did was rebuild the stack from scratch and abandon open source entirely.

[02:09] They're shipping Muse Spark earlier this year as a closed API only model. That created a new problem though, because no sane person would choose Muse Spark over something like Claude or Gemini. So with Wall Street starting to ask questions

[02:22] about the 145 billion of capex being incinerated with nothing to show for it, compound in Kauai that the locals say he stole from them and formed a plan, which eventually led us here to the launch of Muse Glimmer. It's a dense 30 billion

[02:36] parameter model distilled directly from Muse Spark, their big but closed model basically means they had the big model whisper its exact probability until it learned how to fake it. And in Zuck's manifesto he made after the

[02:51] release, he described distillation as an important principle of how the open source ecosystem works, which is ironic because every American lab spent the last year accusing the Chinese of doing this exact same thing to their outputs.

[03:03] So how good is it actually? Well, according to the TMBBs, it clearly beats Gemma 4 and goes bar for bar with Qwen 3.6. But my favorite number is the prompt injection benchmark where attacks against Glimmer succeeded 28% of the

[03:16] time, but they're listing that as a win because with Qwen it's 40% and the way they got it to run on consumer GPUs was interesting. At full precision the model would need over 55 gigs of memory. So their first move was quantization where

[03:29] Meta compressed the weights down to about 4 bits, which shrank it to just under 20 gigs. And the second was speculative decoding, which is basically auto complete for your auto complete. The way it works is a tiny model called

[03:41] DFlash blurts out an entire block of tokens, then the big model reads the whole block in one pass and throws out the bad guesses, which they say led to a 3x speed up on a 5090. But to be honest, the actual model is way less interesting

[03:54] than Zuck's manifesto I mentioned earlier. In it, he argues that the real superintelligence, but a small handful of companies owning it, and that he doesn't understand why anyone who believes AI will end humanity would rush

[04:07] to build it, which is a reasonable point coming from a guy whose company was just fined $567 million last week for being a public nuisance in New Mexico. He also wants Frontier Labs to hand the US government mid-training checkpoints of

[04:20] billion-dollar fund for the towns willing to host his data centers, that ruckus. So, is the redemption arc sincere? Probably not, but an Apache license is an Apache license, and they

[04:33] can't put that genie back in the bottle. Zuck and Wang both say open weights for Muse Spark 1.2 are coming soon, which would mean you could self-host the exact model behind Meta's own coding agent, which is pretty cool. And if you want to

[04:45] can do it with the sponsor of today's video, OpenRouter. It gives you a single API to access every LLM, so you only have one endpoint and one bill to worry about. You can switch between models manually or use their routers to pick

[04:59] one for you based on a combination of factors, like price and accuracy. My startup still hasn't been able to raise a seed round, so I used OpenRouter to rewrite the entire codebase in Rust to attract more investors. Because Rust is

[05:11] a notoriously easy language, I used Pareto Router to route every request to cheaper coding models that could still cargo build, but I eventually ended up the UI, the dedicated image and speech

[05:24] attractive profiles to seed the community and get the matches flowing. OpenRouter is easily the best tool I've seen for discovering new models without subscriptions, and you can try it out for free at the link below. This has

[05:38] watching, and I will see you in the next one.