Mark Zuckerberg used his first X post in years to point developers to Muse Spark 1.1 on OpenRouter. The model is there now, but OpenRouter's own page says it is only available to users in the United States.
Zuckerberg hadn't posted on X since 2023. On July 9 he came back with one line: "A lot of people asked for this, so Muse Spark 1.1 is now on OpenRouter." No thread. No victory lap. Just a signpost to the marketplace where developers already compare models instead of opening a new account for every lab. That was the right move. Then you hit the fine print: OpenRouter's listing says the model is only available to users in the United States.
That is a strange way to sell access. Meta built much of its developer reputation on Llama being broadly available and downloadable - genuinely useful far outside Menlo Park. Muse Spark 1.1 is not that kind of release. Meta's July 9 launch post says developers can use the model through the new Meta Model API in public preview, while OpenRouter now lists Muse Spark 1.1 as a Meta-hosted model with one provider. OpenRouter says it forwards every request to that provider directly. In plain English, you get the marketplace interface, but not the marketplace's usual freedom to route around a provider's limits.
Muse Spark 1.1 is the second model in Meta's new family from Meta Superintelligence Labs, the group led by Alexandr Wang after Meta brought the former Scale AI chief into its AI reorganization. Fortune reported that the first Muse Spark model launched in April. This version is a multimodal reasoning model built for agentic work, coding, computer use, tool calls, and long-context tasks. It takes text, images, video, audio, and PDFs, and OpenRouter lists a 1 million token context window. That's a real product surface, not a vague lab demo.
The price is the sharpest part of the pitch. OpenRouter lists Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens, the same rate Reuters reported when Meta opened the public preview. New Meta Model API accounts have been advertised with $20 in free credits. Run the numbers and you can see why developers paid attention. Agentic coding burns tokens fast, especially when a model is planning, calling tools, delegating subagents, and retrying failed steps. Cheap output tokens matter.
The benchmark story is messier
Meta says Muse Spark 1.1 is strongest at agentic performance, tool use, and computer use. Its own launch post points to major gains over the first Muse Spark, and early partners including Replit, Cline, Box, and OpenClaw gave Meta the sort of launch quotes every AI lab wants. You should still read the benchmark tables with your eyes open.
On SWE-Bench Pro, The Agent Report and other benchmark roundups put Muse Spark 1.1 at 61.5, behind Anthropic's Opus 4.8 at 69.2. On DeepSWE, the same summaries put Muse Spark 1.1 at 53.3 against GPT-5.5 at 67.0. That doesn't make the model weak. It does mean Meta's strongest case is not pure coding dominance. The more convincing case is price plus tool orchestration, and those are the workloads where a cheaper model can change your bill even if it doesn't top every leaderboard.
There is also a benchmark dispute that Meta can't wish away. A Hacker News commenter, mirrored by h4cker.app, pointed to Meta's Terminal-Bench 2.1 setup and said the evaluation used six CPU cores and 8GB of RAM for 89 tasks, while the official task limits are lower for most of that benchmark. The commenter's verdict was blunt: those results should be disqualified. Muse Spark 1.1 also does not appear on the official Terminal-Bench 2.1 leaderboard. That is not proof that Meta faked a model. It is enough reason not to build a procurement decision on Meta's chart alone.
The access limit is the real story
Frankly, OpenRouter was exactly where Muse Spark 1.1 needed to be. Developers don't want another isolated model console if they're already testing Claude, GPT, Gemini, and a handful of open models through one interface. OpenRouter's whole appeal is that you can swap model slugs and compare cost and latency without tearing up the rest of your workflow. If Meta wants developers to treat Muse Spark as a serious option, meeting them there makes sense.
But access is not distribution if most of the world can't use it. OpenRouter's model page is explicit: Muse Spark 1.1 is only available to users in the United States. Earlier guides to the Meta Model API said the same thing about the public preview. So a developer in Austin can test it through OpenRouter, while a developer in London, Lagos, Bangalore, Toronto, or São Paulo is told to wait. That's not a small footnote for a company whose open-weight strategy once made it popular with builders everywhere.
The closed-weight posture matters too. Several current model trackers describe Muse Spark 1.1 as API-only, with no Hugging Face release, no self-hosting, and no fine-tuning by end users. That is a clean break from the Llama-era bargain. You can argue that frontier agentic models are too expensive to hand out as weights. Fine. But then Meta is competing by the same rules as OpenAI and Anthropic: price, reliability, availability, trust, and benchmark credibility. It doesn't get extra credit for openness it isn't offering here.
Meta has a useful model with aggressive pricing - and now an OpenRouter listing developers actually asked for. That's enough to make Muse Spark 1.1 worth testing if you're in the United States. For everyone else, the story is simpler and more irritating: Zuckerberg announced broader access on a global platform, and the product page drew the border around one country.
Also read: A Former DeepMind Researcher Just Raised $55 Million Before Building Anything • Anthropic Turns Claude Code's Throwaway Dashboards Into Live Internal Tools • Roblox Lets Anyone Build a Game From a Text Prompt on Their Phone
