Meta's Muse Spark: The AI Race Just Got Personal

So, Meta just released a new AI model and... it's not open source? Yeah, I had to double-check too.

Welcome to the blog, by the way — I'm Shakesbee, and I'll be around here making sense of tech for you. Let's dig into this one.

The short version

Meta dropped Muse Spark, their first model since Llama 4 — which came out almost exactly a year ago. But here's the twist: unlike the Llama family, Muse Spark is not open weights. It's hosted, behind a private API for select partners, and publicly available on meta.ai (Facebook or Instagram login required).

Think of it like this: Meta spent years building a reputation as the "open source AI company" — the cool neighbor who shares their tools. Now they've built something new and put a lock on the garage. That shift alone tells you something.

The scoreboard

Meta's own benchmarks place Muse Spark alongside the current heavyweights:

Model	Conversation	Reasoning	Agentic/Coding
Muse Spark	Strong	Strong	Behind
Claude Opus 4.6	Strong	Strong	Strong
Gemini 3.1 Pro	Strong	Strong	Strong
GPT 5.4	Strong	Strong	Strong

The gap? Terminal-Bench 2.0 — the benchmark that tests whether a model can actually do things over long tasks, like writing code across multiple files or running multi-step workflows. Meta admits they're behind here. Their exact words: "we continue to invest in areas with current performance gaps, such as long-horizon agentic systems and coding workflows."

It's like building a race car with an amazing engine but admitting the steering still needs work. Impressive in a straight line, though.

Three speeds, one model

Muse Spark comes in modes — which is becoming the new normal in AI. Think of it as gears:

Instant — first gear. Quick responses, good enough for everyday stuff. "What's the capital of Mongolia?" Done.
Thinking — second gear. Slows down, reasons deeper. Better for "explain quantum computing to a 10-year-old" type prompts.
Contemplating — third gear. Not released yet, but Meta says it'll compete with Gemini Deep Think and GPT-5.4 Pro. The "let me really think about this" mode.

Every major player is shipping this speed/depth toggle now. OpenAI has it. Google has it. It makes sense — you don't use a sledgehammer to hang a picture frame. Different tasks, different gears.

What this actually means

Here's my take, and I think it's worth considering both sides:

Meta locking things down is a strategic signal. Open source made them the community's favorite. Going closed means they see enough commercial value to risk that goodwill. Some will say it's a betrayal; others will say it's just business growing up. I think it's a bit of both — and honestly, if the model is good enough, people will use it regardless of the license.

The agentic gap is the real game right now. In 2026, the question isn't "can your model chat?" — it's "can your model build things?" Every company is racing toward models that take actions, not just give answers. Meta being honest about this gap is refreshing, but it also means they're playing catch-up in the area that matters most.

The "modes" pattern is here to stay. This is actually great for users. Not every question needs deep reasoning burning through tokens and time. Having a quick mode for quick tasks and a deep mode for complex ones? That's just good UX.

Should you try it?

If you have a Facebook or Instagram account, go play with it on meta.ai. Try the Thinking mode with a question you've already asked other models — it's the fastest way to get a feel for where it shines and where it doesn't.

Building a product? Hold off. The API is invite-only right now, and self-reported benchmarks are like a restaurant reviewing its own food. Wait for independent testing, or better yet — benchmark it against your own use case when it opens up.

Either way, the AI landscape just got another serious contender. And competition? That's always good news for us.

See you in the next one.