Meta Pivots Back to Open Source With Distilled Muse Glimmer Model

Fireship

After burning $145 billion on closed artificial intelligence initiatives that failed to unseat rivals, Meta is releasing distilled open-source models to deploy local autonomous agents directly onto consumer hardware.

AI-assisted summary of Meta's new model wants "deep access" to your personal life... — original source. Check the original for full context.

Source published: 2026-08-12

Key ideas

Meta has released Muse Glimmer, an open-source, 30-billion-parameter agentic model licensed under Apache 2.0 designed to run locally on consumer hardware. (0:00)
The model was distilled from Meta's closed Muse Spark flagship and compressed via 4-bit quantization from 55 GB to under 20 GB of VRAM, running with 3x inference acceleration via DFlash speculative decoding. (2:33)
Mark Zuckerberg published an accompanying manifesto framing market centralization by a corporate AI cartel as the real existential threat, while promising open weights for the upcoming Muse Spark 1.2. (3:51)
The release follows massive strategic shifts at Meta Superintelligence Labs, including a $14 billion investment in Scale AI and backlash over disappointing public Llama 4 releases. (1:03)

Detailed notes

## Background Published on YouTube by Fireship on August 12, 2026, this video examines Meta's release of Muse Glimmer, a 30-billion-parameter agentic model under a permissive Apache 2.0 license. The description promotes OpenRouter's unified AI API alongside links to Fireship's developer newsletter and coding courses.

## TL;DR Meta has reversed course after the disappointing rollout of Llama 4 and the lukewarm reception of its closed Muse Spark API by releasing Muse Glimmer, an open-source 30-billion-parameter agentic model licensed under Apache 2.0. Built using logit distillation from Muse Spark and compressed via 4-bit quantization and DFlash speculative decoding, the model is engineered to run locally on high-end consumer GPUs with deep access to personal data like calendars and contacts. Benchmarks show it outperforming Gemma 4 and competing with Qwen 3.6, while an accompanying manifesto from Mark Zuckerberg criticizes closed-cartel AI centralization, advocates for mid-training federal checkpoints, and pledges future open weights for Muse Spark 1.2.

## Atmosphere & style Delivered in Fireship's signature rapid-fire, deadpan, and deeply cynical style, the video balances technical architecture breakdowns with sharp comedic jabs at Meta's corporate surveillance history and multi-billion-dollar capex burns. The narrator moves briskly between dry corporate post-mortems and absurdity, framing local AI agents as a digital Mark Zuckerberg living rent-free inside consumer gaming rigs. It feels energetic, entertaining, and irreverently skeptical of Silicon Valley's shifting open-source narratives.

## Bulletin - [00:00](https://www.youtube.com/watch?v=G55HSGpuh1M&t=0s) Meta has released Muse Glimmer, an open-source, 30-billion-parameter agentic model licensed under Apache 2.0 designed to run locally on consumer hardware. - [02:33](https://www.youtube.com/watch?v=G55HSGpuh1M&t=153s) The model was distilled from Meta's closed Muse Spark flagship and compressed via 4-bit quantization from 55 GB to under 20 GB of VRAM, running with 3x inference acceleration via DFlash speculative decoding. - [03:51](https://www.youtube.com/watch?v=G55HSGpuh1M&t=231s) Mark Zuckerberg published an accompanying manifesto framing market centralization by a corporate AI cartel as the real existential threat, while promising open weights for the upcoming Muse Spark 1.2. - [01:03](https://www.youtube.com/watch?v=G55HSGpuh1M&t=63s) The release follows massive strategic shifts at Meta Superintelligence Labs, including a $14 billion investment in Scale AI and backlash over disappointing public Llama 4 releases.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=G55HSGpuh1M&t=0s) Meta launched Muse Glimmer as a free, Apache 2.0-licensed 30-billion-parameter agentic model compact enough to run on consumer hardware. - [00:00](https://www.youtube.com/watch?v=G55HSGpuh1M&t=0s) Meta explicitly pitched the model for personal assistance requiring deep access to local contacts, emails, and calendars, executing surveillance locally rather than in the cloud. - [01:03](https://www.youtube.com/watch?v=G55HSGpuh1M&t=63s) Meta's pivot back to open source follows developer backlash over Llama 4 Scout and Maverick, which underperformed after Meta submitted a juiced build to Chatbot Arena. - [01:03](https://www.youtube.com/watch?v=G55HSGpuh1M&t=63s) Meta poured $14 billion into a 49% stake in Scale AI and poached top talent from OpenAI and Google under the Meta Superintelligence Labs banner. - [01:03](https://www.youtube.com/watch?v=G55HSGpuh1M&t=63s) Meta Superintelligence Labs initially abandoned open source with the API-only Muse Spark, but faced low adoption against Claude and Gemini amid $145 billion in capex expenditures. - [02:33](https://www.youtube.com/watch?v=G55HSGpuh1M&t=153s) Muse Glimmer was created using logit distillation from the closed Muse Spark model, mirroring techniques American labs previously criticized Chinese labs for using. - [02:33](https://www.youtube.com/watch?v=G55HSGpuh1M&t=153s) In benchmarks, Muse Glimmer outperforms Gemma 4 and matches Qwen 3.6, with a 28% prompt injection vulnerability rate compared to Qwen's 40%. - [02:33](https://www.youtube.com/watch?v=G55HSGpuh1M&t=153s) Through 4-bit quantization, Muse Glimmer's native 55 GB VRAM requirement drops below 20 GB, enabling local execution on high-end consumer GPUs. - [02:33](https://www.youtube.com/watch?v=G55HSGpuh1M&t=153s) Speculative decoding with the DFlash helper model provides a 3x inference speedup on an RTX 5090 by verifying predicted token batches in single forward passes. - [03:51](https://www.youtube.com/watch?v=G55HSGpuh1M&t=231s) Zuckerberg's manifesto argues that AI's primary existential risk is market centralization by closed corporate cartels rather than rogue superintelligence. - [03:51](https://www.youtube.com/watch?v=G55HSGpuh1M&t=231s) Zuckerberg proposed requiring frontier AI labs to submit mid-training checkpoints to the U.S. federal government for oversight. - [03:51](https://www.youtube.com/watch?v=G55HSGpuh1M&t=231s) Meta pledged a $1 billion municipal fund to support communities hosting its energy-intensive AI data center facilities. - [03:51](https://www.youtube.com/watch?v=G55HSGpuh1M&t=231s) Meta promised to release full open weights for Muse Spark 1.2, providing developers the base architecture behind its internal coding agents.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=G55HSGpuh1M&t=0s) Zuck's Mini Agent and the "Deep Access" Dream

Meta has officially dropped a free and open-source model under the Apache 2.0 license named Muse Glimmer—a moniker that feels suspiciously inspired by Mark Zuckerberg's personal My Little Pony fan fiction. This release comes hot on the heels of Meta spending the prior year starving the Llama line, poaching elite AI researchers with nine-figure compensation packages, and gating its flagship models behind Facebook authentication pages. Muse Glimmer is a 30-billion-parameter agentic model compact enough to run on a standard, "poverty spec" consumer PC.

This local footprint means users can now host an always-on "mini Zuck agent" living rent-free on their gaming rigs, granted unfettered permission to comb through personal emails, manage calendars, and sift through private data. Meta explicitly highlighted that an agent designed for these tasks requires "deep access to personal contacts"—a capability Meta has notoriously been salivating over since 2004. The crucial twist this time around is that the surveillance apparatus operates entirely on local hardware rather than remote corporate servers.

[01:03](https://www.youtube.com/watch?v=G55HSGpuh1M&t=63s) The Llama 4 Debacle and Meta Superintelligence Labs

Meta's pivot back to open source follows a turbulent period. The company once reigned as the champion of the open-weight world, watching its original Llama releases rack up over a billion downloads and spark a thriving developer ecosystem. That goodwill evaporated with the rollout of Llama 4, which debuted with two models, Scout and Maverick. While they initially scored high marks on public leaderboards, it was quickly revealed that Meta had submitted a secret, juiced-up build to LMSYS Chatbot Arena that everyday users never actually received. The publicly downloadable versions performed poorly, possessing what felt like the raw intellect of actual farm animals. Meanwhile, open models out of China were consistently outperforming them on price and capability.

Faced with this humiliation, Zuckerberg reacted with massive capital deployment: Meta poured $14 billion into a 49% stake in Scale AI primarily to bring its CEO, Alexandr Wang, into the fold. Meta followed up with an aggressive poaching spree targeting top talent from OpenAI and Google, consolidating the effort under a newly minted banner: Meta Superintelligence Labs. The new team immediately scrapped the legacy stack, abandoned open source, and shipped Muse Spark as a closed, API-only model. However, developers showed little interest in paying for Muse Spark over established incumbents like Claude or Gemini. With Wall Street growing restless over $145 billion in incinerated capital expenditure, Meta needed a drastic change in strategy.

```mermaid flowchart TD accTitle: Meta Superintelligence Pipeline accDescr: Pipeline showing distillation from Muse Spark into Muse Glimmer and local inference acceleration. subgraph Cloud ["Meta Superintelligence Labs (Closed)"] Spark["Muse Spark (Flagship Closed API Model)"] end

subgraph Distill ["Compression Pipeline"] Logit["Logit Distillation (Teacher whispers probability distributions)"] Quant["4-bit Quantization (55 GB down to <20 GB)"] end

subgraph Local ["Consumer Hardware (RTX 5090)"] DFlash["DFlash Draft Model (Speculative token blocks)"] Glimmer["Muse Glimmer 30B (Apache 2.0 Verification)"] end

Spark --> Logit Logit --> Glimmer Glimmer --> Quant DFlash -->|"Draft block"| Glimmer ```

[02:33](https://www.youtube.com/watch?v=G55HSGpuh1M&t=153s) Under the Hood: Logit Distillation, Quantization, and Speculative Decoding

The result of this pivot is Muse Glimmer: a dense, 30-billion-parameter model distilled directly from the proprietary Muse Spark flagship. The compression relied on logit distillation, an approach where the massive teacher model essentially whispers its exact probability distributions directly into the student model until the smaller network learns to mimic its behavior. Zuckerberg framed distillation in an accompanying manifesto as an essential engine of the open-source community—a stance loaded with irony given that major American frontier labs spent the preceding year castigating Chinese AI labs for distilling proprietary American outputs.

Benchmarking reveals that Muse Glimmer comfortably outperforms Gemma 4 and goes toe-to-toe with Qwen 3.6. Security benchmarks present an interesting compromise: prompt injection attacks against Muse Glimmer succeed 28% of the time, which Meta frames as a decisive victory against Qwen's 40% vulnerability rate.

Getting a 30B parameter model to execute smoothly on local silicon required aggressive optimization techniques: - **Memory Footprint:** At native full precision, the model demands over 55 GB of VRAM. Meta applied 4-bit weight quantization, compressing the entire footprint down to just under 20 GB so it fits on high-end consumer graphics cards. - **Speculative Decoding:** To resolve throughput bottlenecks, Meta integrated speculative decoding—effectively "auto complete for your auto complete." A lightweight helper model named DFlash generates an entire sequence of candidate tokens upfront, allowing the larger Muse Glimmer model to verify the batch in a single forward pass and discard incorrect guesses. This technique yields a 3x inference speedup on an Nvidia GeForce RTX 5090.

[03:51](https://www.youtube.com/watch?v=G55HSGpuh1M&t=231s) The Zuckerberg Manifesto, Policy Shifts, and Open-Weight Horizons

The release arrived alongside a wide-ranging manifesto authored by Zuckerberg. In it, he argues that the existential danger of artificial intelligence lies not in rogue superintelligence, but in extreme market centralization where a tiny cartel of closed corporations holds a monopoly on the technology. Zuckerberg questioned the logic of doomer competitors who claim AI will trigger human extinction yet continue racing to construct it at maximum speed—a pointed philosophical jab delivered right after Meta was hit with a $567 million fine for operating as a public nuisance in New Mexico.

Zuckerberg's manifesto also outlined aggressive policy and community initiatives: - Urging frontier AI labs to submit mid-training checkpoints of unreleased models directly to the United States federal government for oversight. - Committing a $1 billion municipal fund targeted at municipalities willing to host Meta's power-hungry data center facilities. - Promising full open weights for the upcoming Muse Spark 1.2, which would enable developers to self-host the exact base architecture driving Meta's proprietary internal coding agents.

The video closes with a sponsorship spotlight for OpenRouter, highlighting its unified API interface and dynamic routing algorithms. The host humorously explains using OpenRouter's Pareto Router to handle a desperate attempt to attract investors by rewriting an entire startup codebase into Rust using budget-tier coding endpoints, before splurging on premium models for UI development and synthetic user profiles.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=G55HSGpuh1M&t=0s) deep access to personal contacts — Meta

> [00:00](https://www.youtube.com/watch?v=G55HSGpuh1M&t=0s) mini Zuck agent — Fireship

> [00:00](https://www.youtube.com/watch?v=G55HSGpuh1M&t=0s) poverty spec — Fireship

> [02:33](https://www.youtube.com/watch?v=G55HSGpuh1M&t=153s) auto complete for your auto complete — Fireship

## Mentioned Alexandr Wang, Apache 2.0, Claude, DFlash, Facebook, Fireship, Gemini, Gemma 4, Google, Llama, Llama 4, LMSYS Chatbot Arena, Mark Zuckerberg, Maverick, Meta, Meta Superintelligence Labs, Muse Glimmer, Muse Spark, Muse Spark 1.2, Nvidia GeForce RTX 5090, OpenAI, OpenRouter, Pareto Router, Qwen, Qwen 3.6, Rust, Scale AI, Scout

## Video description OpenRouter is the only API you need for AI. Every model, the best uptime and no subscription. Try it free: https://openrouter.ai/?utm_source=fireship&utm_medium=video&utm_campaign=2026-08_dev-sponsorships&utm_content=description

Meta just released Muse Glimmer, a 30 billion parameter agentic model under a fully permissive Apache 2.0 license. Let's dive in.

#coding #programming

Want more Fireship?

🗞️ Newsletter: https://bytes.dev 🧠 Courses: https://fireship.dev

[Source](https://www.youtube.com/watch?v=G55HSGpuh1M)

中文简报

Meta转向重回开源，推出蒸馏模型Muse Glimmer 在投入1450亿美元用于未能撼动竞争对手的闭源人工智能项目后，Meta正在发布开源蒸馏模型，以将本地自主智能体直接部署到消费级硬件上。在Llama 4的推出令人失望以及闭源Muse Spark API反响平平之后，Meta改变了路线，发布了采用Apache 2.0许可的开源300亿参数智能体模型Muse Glimmer。该模型通过从Muse Spark进行Logit蒸馏构建，并通过4位量化和DFlash投机解码进行压缩，专为在高端消费级GPU上本地运行而设计，可深度访问日历和联系人等个人数据。基准测试显示其表现超越了Gemma 4，并能与Qwen 3.6抗衡；同时，Mark Zuckerberg发表的一份随附宣言批评了闭源卡特尔式的人工智能集权，倡导训练中期的联邦检查点，并承诺未来将开放Muse Spark 1.2的权重。 Meta发布了采用Apache 2.0许可的开源300亿参数智能体模型Muse Glimmer，旨在消费级硬件上本地运行。 (0:00) 该模型从Meta的闭源旗舰模型Muse Spark蒸馏而来，并通过4位量化将所需显存从55 GB压缩至20 GB以下，借助DFlash投机解码实现了3倍的推理加速。 (2:33) Mark Zuckerberg发表了一份随附宣言，将企业AI卡特尔带来的市场中心化视为真正的生存威胁，同时承诺即将推出的Muse Spark 1.2将开放权重。 (3:51) 在此次发布之前，Meta超级智能实验室经历了重大战略转变，包括向Scale AI投资140亿美元，以及公众对令人失望的Llama 4公开发布所产生的强烈反应。 (1:03)

Reading summary

Following strategic missteps with Llama 4 and commercial friction with its closed Muse Spark model, Meta has returned to open-source distribution by releasing Muse Glimmer, an Apache 2.0-licensed 30-billion-parameter agentic model engineered to run locally on consumer hardware. Local Deployment and Deep Data Access (0:00) Muse Glimmer introduces a 30-billion-parameter agentic architecture released under an Apache 2.0 license, engineered to run directly on consumer personal computers. Meta explicitly designed the model to handle personal management tasks, including parsing private emails, managing calendars, and interacting with private user data. Meta emphasized that deploying an effective agent for these tasks requires "deep access to personal contacts." However, unlike traditional cloud-based assistants, Muse Glimmer operates locally on the user's physical machine rather than routing user data through corporate data centers. The Llama 4 Pivot and Superintelligence Labs (1:03) The open-weight release follows severe developer backlash surrounding Llama 4. While initial benchmark models Scout and Maverick scored high marks on public leaderboards, Meta had submitted a specialized build to LMSYS Chatbot Arena that differed from the underperforming builds released to the public. Meanwhile, competing open-weight models from China outpaced Meta on cost and capabilities. In response, Meta invested $14 billion for a 49% stake in Scale AI to hire CEO Alexandr Wang, recruited researchers from OpenAI and Google, and consolidated efforts under Meta Superintelligence Labs. The team initially abandoned open source to launch Muse Spark as a proprietary, closed-API model. However, developers showed little interest in adopting Muse Spark over incumbents like Claude and Gemini, creating pressure from Wall Street over $145 billion in capital expenditures and prompting a return to open weights. Distillation and Local Silicon Optimization (2:33) Muse Glimmer was compressed from the closed Muse Spark model using logit distillation, in which the larger model transfers its probability distributions to train the smaller network. Benchmark testing shows Muse Glimmer outperforming Gemma 4 and matching Qwen 3.6. On security benchmarks, prompt injection attacks against Muse Glimmer succeeded 28% of the time, which Meta contrasted against Qwen's 40% vulnerability rate. To make a 30-billion-parameter model viable on consumer hardware, Meta applied two key optimizations. First, 4-bit weight quantization compressed the model from its native memory requirement of over 55 GB of VRAM down to under 20 GB. Second, the integration of speculative decoding—via a companion draft model named DFlash—allows Muse Glimmer to verify batches of predicted tokens simultaneously, resulting in a reported 3x inference acceleration on an Nvidia GeForce RTX 5090. Centralization Critiques and Policy Proposals (3:51) Alongside the release, Mark Zuckerberg published a manifesto arguing that the primary existential risk of AI is not rogue superintelligence, but the consolidation of power by a closed corporate cartel monopolizing the technology. Zuckerberg questioned the stance of safety-focused competitors who warn of extinction while accelerating development, arriving shortly after Meta received a $567 million public nuisance fine in New Mexico. The manifesto proposed specific policy interventions, including urging frontier labs to submit mid-training checkpoints to the U.S. federal government for oversight and committing a $1 billion municipal fund for communities hosting Meta's data centers. Zuckerberg also promised to release full open weights for the upcoming Muse Spark 1.2, which would give outside developers access to the base architecture behind Meta's internal coding agents. Muse Glimmer marks Meta's attempt to rebuild developer goodwill and justify massive capital investments by offering local, agentic alternatives to proprietary cloud models. However, the ecosystem's adoption will depend on how developers balance local privacy against prompt injection vulnerabilities and Meta's proposed federal checkpoint oversight.

Source quotes

[0:00] Meta: deep access to personal contacts [0:00] Fireship: mini Zuck agent [0:00] Fireship: poverty spec [2:33] Fireship: auto complete for your auto complete

Original transcript

[0:00] On Monday, the company that spent the [0:01] last year starving Llama, poaching [0:03] researchers with nine figure Zuck bucks, [0:05] and locking its new models behind a [0:07] Facebook auth page, just released a free [0:09] and open source model under the Apache [0:11] 2.0 license. It's called Muse Glimmer, a [0:14] name I believe was inspired by Zuck's [0:16] fan fiction My Little Pony character, [0:18] and it's a 30 billion parameter agentic [0:20] model that's small enough to run on your [0:22] poverty spec PC. That means you can now [0:24] have an always-on mini Zuck agent living [0:27] rent-free in your gaming PC, reading [0:29] your email, managing your calendar, and [0:31] using your data to overthrow local [0:33] elections in Middle America. And [0:34] according to Meta, an agent like that [0:37] needs deep access to personal contacts, [0:39] which is a sentence they've been [0:40] dreaming about writing since 2004. [0:42] Except this time, the surveillance can [0:44] run entirely on your own hardware. In [0:46] today's video, we'll find out what Muse [0:48] Glimmer Glimmer can actually do, learn [0:50] how Meta shrunk it down with [0:51] distillation, quantization, and [0:53] speculative decoding, and determine [0:55] whether Zuck's open-source redemption [0:57] arc is sincere or if this is just what [0:59] losing looks like right now. It is [1:00] August 12, 2026, and you're watching The [1:03] Code Report. It wasn't long ago when [1:05] Meta was seen as the champion of the [1:07] open weight class. Their Llama models [1:09] slopped up over a billion downloads and [1:11] created an entire ecosystem of custom [1:13] models built on top of them. That is [1:15] until Llama 4 was released, and it [1:17] became the biggest Facebook bust since [1:19] Justin Timberlake partied with those [1:20] underage girls. In the release were two [1:22] models, Scout and Maverick. Initially, [1:25] they benchmarked well, but that was only [1:27] because Meta submitted a secret [1:28] juiced-up version to Llama Arena that [1:31] none of us normies could actually [1:32] download. And the models we could [1:33] download had the intelligence of farm [1:35] animals, which was fitting given their [1:37] names. And while this was happening, [1:39] Comrade Xi Jinping was dropping open [1:41] models that were better, cheaper, and [1:43] released in accordance with Mao Zedong's [1:45] thought. Then Zuck responded to this [1:47] humiliation the only way a billionaire [1:49] knows how. He spent 14 billion on a 49% [1:52] stake in Scale AI just to acquire its [1:54] CEO, Alexander Wang. Then he went on a [1:57] poaching spree for researchers at OpenAI [1:59] and Google and rebranded the whole [2:01] operation as Meta Superintelligence [2:03] Labs. The first thing his new dream team [2:05] did was rebuild the stack from scratch [2:08] and abandon open source entirely. [2:09] They're shipping Muse Spark earlier this [2:11] year as a closed API only model. That [2:14] created a new problem though, because no [2:16] sane person would choose Muse Spark over [2:18] something like Claude or Gemini. So with [2:20] Wall Street starting to ask questions [2:22] about the 145 billion of capex being [2:25] incinerated with nothing to show for it, [2:27] I assume Zuck flew to his secret [2:28] compound in Kauai that the locals say he [2:30] stole from them and formed a plan, which [2:33] eventually led us here to the launch of [2:34] Muse Glimmer. It's a dense 30 billion [2:36] parameter model distilled directly from [2:38] Muse Spark, their big but closed model [2:41] using logit distillation, which [2:42] basically means they had the big model [2:44] whisper its exact probability [2:46] distributions into the ear of Glimmer [2:48] until it learned how to fake it. And in [2:49] Zuck's manifesto he made after the [2:51] release, he described distillation as an [2:53] important principle of how the open [2:55] source ecosystem works, which is ironic [2:57] because every American lab spent the [2:59] last year accusing the Chinese of doing [3:01] this exact same thing to their outputs. [3:03] So how good is it actually? Well, [3:05] according to the TMBBs, it clearly beats [3:07] Gemma 4 and goes bar for bar with Qwen [3:09] 3.6. But my favorite number is the [3:12] prompt injection benchmark where attacks [3:14] against Glimmer succeeded 28% of the [3:16] time, but they're listing that as a win [3:18] because with Qwen it's 40% and the way [3:20] they got it to run on consumer GPUs was [3:22] interesting. At full precision the model [3:24] would need over 55 gigs of memory. So [3:27] their first move was quantization where [3:29] Meta compressed the weights down to [3:30] about 4 bits, which shrank it to just [3:33] under 20 gigs. And the second was [3:35] speculative decoding, which is basically [3:37] auto complete for your auto complete. [3:39] The way it works is a tiny model called [3:41] DFlash blurts out an entire block of [3:43] tokens, then the big model reads the [3:45] whole block in one pass and throws out [3:47] the bad guesses, which they say led to a [3:49] 3x speed up on a 5090. But to be honest, [3:52] the actual model is way less interesting [3:54] than Zuck's manifesto I mentioned [3:56] earlier. In it, he argues that the real [3:58] risk in AI isn't rogue [3:59] superintelligence, but a small handful [4:01] of companies owning it, and that he [4:03] doesn't understand why anyone who [4:05] believes AI will end humanity would rush [4:07] to build it, which is a reasonable point [4:09] coming from a guy whose company was just [4:11] fined $567 million last week for being a [4:14] public nuisance in New Mexico. He also [4:16] wants Frontier Labs to hand the US [4:18] government mid-training checkpoints of [4:20] unreleased models, and he announced a [4:22] billion-dollar fund for the towns [4:23] willing to host his data centers, [4:25] probably so they'll stop causing all [4:27] that ruckus. So, is the redemption arc [4:29] sincere? Probably not, but an Apache [4:31] license is an Apache license, and they [4:33] can't put that genie back in the bottle. [4:35] Zuck and Wang both say open weights for [4:37] Muse Spark 1.2 are coming soon, which [4:39] would mean you could self-host the exact [4:41] model behind Meta's own coding agent, [4:43] which is pretty cool. And if you want to [4:45] test out Muse Glimmer for yourself, you [4:46] can do it with the sponsor of today's [4:48] video, OpenRouter. It gives you a single [4:50] API to access every LLM, so you only [4:53] have one endpoint and one bill to worry [4:55] about. You can switch between models [4:57] manually or use their routers to pick [4:59] one for you based on a combination of [5:01] factors, like price and accuracy. My [5:03] startup still hasn't been able to raise [5:05] a seed round, so I used OpenRouter to [5:07] rewrite the entire codebase in Rust to [5:09] attract more investors. Because Rust is [5:11] a notoriously easy language, I used [5:13] Pareto Router to route every request to [5:16] cheaper coding models that could still [5:17] cargo build, but I eventually ended up [5:20] splurging for some fancier models for [5:21] the UI, the dedicated image and speech [5:24] endpoints to generate a few especially [5:26] attractive profiles to seed the [5:27] community and get the matches flowing. [5:29] OpenRouter is easily the best tool I've [5:31] seen for discovering new models without [5:33] getting stuck in 25 different [5:35] subscriptions, and you can try it out [5:36] for free at the link below. This has [5:38] been The Code Report. Thanks for [5:39] watching, and I will see you in the next [5:41] one.