AI at Meta

Simon Willison

Pietro Schirano

<p>Meta has released Muse Spark, the first model from its newly formed Meta Superintelligence Labs and the first model it has built from scratch rather than iterating on Llama. It’s been nine months in the making under chief AI officer Alexandr Wang, and it’s available now at <a href="http://meta.ai" target="_blank" rel="noopener nofollow">meta.ai</a> — though not as open weights, which is a notable break from Meta’s usual approach.</p>
<h2>What Muse Spark actually is</h2>
<p>It’s a natively multimodal reasoning model with tool use, visual chain-of-thought, and multi-agent orchestration baked in. The headline feature is <strong>Contemplating mode</strong>, which runs multiple sub-agents in parallel on harder tasks. Meta claims this gets Muse Spark to 58% on Humanity’s Last Exam. There’s also a standard Thinking mode and a faster Instant mode.</p>
<p>One interesting technical claim: a training technique called “thought compression” that Meta says achieves equivalent performance to Llama 4 Maverick using more than 10x less compute during pretraining. That’s a significant efficiency gain if it holds up.</p>
<h2>The tools wired into <a href="http://meta.ai" target="_blank" rel="noopener nofollow">meta.ai</a></h2>
<p>Simon Willison catalogued 16 tools available through the <a href="http://meta.ai" target="_blank" rel="noopener nofollow">meta.ai</a> chat interface. The notable ones:</p>
<ul>
<li>Web browsing</li>
<li>Meta social content search (searches across Facebook, Instagram, etc.)</li>
<li>Python code execution (Code Interpreter-style)</li>
<li>HTML and SVG artifact creation</li>
<li>Image generation</li>
<li>Sub-agent spawning for parallel tasks</li>
<li>A visual grounding tool — likely powered by Segment Anything — that can identify, locate, and count objects in images</li>
</ul>
<p>In practice, people are already finding it useful for image-to-code conversion and generating SVG artwork. The SVG output in particular has gotten some attention.</p>
<h2>How the benchmarks actually look</h2>
<p>Mixed, honestly. Muse Spark ranks fourth on the Artificial Analysis Intelligence Index, behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. On abstract reasoning (ARC AGI 2), it scores 42.5 compared to rivals sitting around 76 — a meaningful gap.</p>
<p>Where it does well: HealthBench Hard (42.8%, the top score) and chart understanding. The health performance makes sense given Meta says it developed those capabilities with input from over 1,000 physicians.</p>
<h2>The closed-source question</h2>
<p>This is the part that will bother some people. Muse Spark is closed source — Meta frames this as temporary and says it “hopes” to open-source future versions, but there’s no commitment or timeline. For a company that built its AI reputation on releasing Llama weights publicly, this is a real shift.</p>
<p>The model is available to all users on <a href="http://meta.ai" target="_blank" rel="noopener nofollow">meta.ai</a> for testing. A private API preview is opening to select partners. It now powers Meta AI across the company’s platforms, with dedicated shopping and health modes that lean on Meta’s platform data.</p>
<p>This release is also the first product output from Meta’s $14.3 billion Scale AI investment, part of a broader $115–135 billion 2026 capital expenditure plan. There’s a lot riding on whether Muse Spark can close the gap with the models currently ahead of it on the benchmarks that matter most.</p>


Collections

Meta has released Muse Spark, its first model built from scratch under Meta Superintelligence Labs, departing from its Llama lineage. It is a natively multimodal reasoning model featuring Contemplating mode (parallel sub-agents), Thinking mode, and Instant mode, with 16 tools integrated into meta.ai including web browsing, Python code execution, image generation, and a visual grounding tool. A training technique called 'thought compression' claims 10x compute efficiency gains over Llama 4 Maverick. Benchmark results are mixed: it ranks fourth on the Artificial Analysis Intelligence Index behind Gemini, GPT-5, and Claude, with a notable gap on ARC AGI 2 (42.5 vs ~76 for rivals), but leads on HealthBench Hard. Notably, Muse Spark is closed source — a significant departure from Meta's open-weights tradition — with no firm commitment to open-sourcing future versions. It is available on meta.ai and powering Meta AI across the company's platforms, backed by Meta's $14.3B Scale AI investment.

Meta releases Muse Spark, its first natively multimodal reasoning model