Meta is back in AI with a new Spark

Mark Zuckerberg’s Grand Return to X

Zuckerberg came back to Elon's app to sell his new model which is very different from the Llama models he previously released. This is after the open source Llama models fell behind. It is Meta’s big attempt to get back into the top-tier AI race.

You see the thing is Meta used to push Open Source models which are models you could download the weights and run them yourself. Some people even said Meta is the pioneer of Open Source though most disagree because of how China has done better when it comes to open source frontier models like Kimi 3 and GLM 5.2. By frontier I mean these are kind of models that are competing against the best from Open AI and Anthropic in certain BenchMarks.

Getting back to Meta so after losing the open source model competition against China Mark is now joining the Open AI, Google, Anthropic, and XAI in the closed models AI race with a model of his own he calls Muse. He created the Meta Superintelligence Labs in June 2025 and brought in Alexandr Wang, a young founder of Scale AI, the big data-labeling company as Chief AI Officer to lead it. Meta also put a huge amount of money into Scale AI as part of the deal ($14.3 billion for a 49% stake).

On July 9, 2026, Mark posted a thread not on Threads but X announcing Muse Spark 1.1. He followed up explaining its strengths: agentic performance, tool use, computer use, a 1-million-token context window, the ability to spin up parallel sub-agents, and training to control computer interfaces across desktop, mobile, or browser. The focus, he said, was delivering strong agentic and multimodal models at very low cost.

The tech world noticed immediately. Nikita Bier, then still closely tied to X product and later reflecting on his time as head of product, welcomed him with characteristic dryness: “Hello Mark, would you like me to enable Creator Monetization on your account.”

Elon Musk simply replied “Jinx” — a one-word post that drew laughs and speculation about timing, competition, or pure meme energy.

Elon meant the classic “jinx” game version the one where two people say (or do) the same thing at the same time and one calls “jinx.”

A week later Zuckerberg posted that Muse Spark 1.1 was now on OpenRouter after many requests. When a developer asked what separated it from the rest of the field, Zuckerberg replied simply: “High intelligence at a pretty aggressive price imo.”

That pricing stance has become a recurring theme.

Open weights, closed models, and the same page

Even while shipping closed models, Zuckerberg has kept the open-source conversation alive. When Microsoft CEO Satya Nadella posted about open-weight models being essential to a healthy ecosystem and American competitiveness, Zuckerberg quote-posted: “Open source is a positive and important force for both empowering people and preventing centralization. Proud to support this.” On this issue, at least, Meta, Microsoft, and others (including Elon at times) have found common ground.

Meta’s current strategy is clear: ship capable closed models that can be tightly integrated into Facebook, Instagram, WhatsApp, Ray-Ban glasses, and monetized through a real public API, while still signaling support for the broader open ecosystem.

So now, What exactly is Muse Spark?

Muse Spark is Meta’s new flagship AI model family. It is a proprietary closed multimodal reasoning model built by Meta Superintelligence Labs the new AI unit led by Alexandr Wang. Unlike the earlier open Llama models, you cannot download Muse Spark’s weights you use it only through Meta’s products or their paid API.

But looking at real developer reactions over the last month, here’s what people keep coming back to.

1. The pricing is the real moat

This is the thing that keeps surprising people.

• Standard API: $1.25 input / $4.25 output per million tokens

• Contributor tier: $0.10 input / $0.20 output (Meta can train on your data)

Multiple developers have called the contributor pricing “INDECENTLY CHEAP” or “Intelligence by the meter.” One person noted it’s cheaper than DeepSeek-v4-Flash while performing closer to GPT-5.5-class models. Another said Muse Spark 1.2 hit top 5 on the Vals Index at roughly $0.69 per test — 3x cheaper than Kimi and 10x+ cheaper than Opus / Fable / Sol-class models.

A lot of people are treating it as the “default cheap but still strong” model they throw at high-volume or long-running work.

2. It’s built for actual work, not just chatting

Users keep highlighting the agentic features more than pure intelligence scores:

• Persistent background agents that stay alive the entire session (so they don’t lose context every time)

• Ability to fan out parallel sub-agents into isolated worktrees (no collisions)

• Full local event log — if it crashes, it resumes exactly where it left off

• Comfortable running jobs that take minutes or even hours (one internal example was 24 hours and 1,000+ tool calls)

One developer put it simply:

“The part that caught my eye wasn’t even the benchmarks. It was the fact that Muse Spark 1.2 appears comfortable working through long engineering jobs that take minutes instead of seconds. That’s where real enterprise value gets created.”

3. Muse Code

Muse Code (the terminal agent powered by Spark 1.2) is what a lot of people are actually testing. Feedback so far:

• Good at planning + writing + validating across large repos

• Feels more “system-oriented” than just another coding model

• Some users who were frustrated with Sol or Opus limits said they were “very pleasantly surprised” after switching for a few hours

• Others treat it as an excellent fast/cheap sub-agent inside a larger agent harness rather than the single “god model”

But of Course there are Honest trade-offs people mention

• Not always the absolute smartest model for complex creative tasks (some game-generation tests still preferred GPT-5.6 Soul or Qwen)

• The ultra-cheap pricing comes with the data-sharing trade-off on the contributor tier

• Still early, so the “vibe” and reliability are still being stress-tested by the community

The Llama era was about open weights and broad adoption. The Muse era, so far, is about personal superintelligence, aggressive efficiency, and agents that do real work. Whether that combination is enough to close the remaining gaps is the question the next few releases will answer. For now, Zuckerberg is posting on X again, the models are shipping, and Meta is once more a serious participant in the AI race.