Metas Muse Spark arrives a year late to a changed race

What Muse Spark is

Muse Spark is a small, fast model. Meta described it as capable enough to reason through complex questions in science, maths, and health. It will initially be available on the Meta AI app and website. Within weeks, it will replace the existing Llama models powering chatbots across WhatsApp, Instagram, Facebook, and Meta's smart glasses.

Unlike previous Llama releases, Muse Spark is not open-source. Meta shared only a "private preview" with unnamed partners. The company did not disclose the model's size. Bigger versions are in development, and Alex Wang has said Meta plans to release at least some of them openly.

The benchmarks tell half the story

Independent evaluations showed Muse Spark catching up with top models from Google, OpenAI, and Anthropic on language and visual understanding. It fell behind on coding and abstract reasoning.

Mark Zuckerberg had set expectations for this. In January, he told investors the first models from the team would be good but, more importantly, would show the trajectory Meta is on. Wang acknowledged rough edges in a series of social media posts.

For anyone tracking AI purely on benchmark scores, fourth place after spending $14.3 billion looks like a poor return. But benchmarks measure one kind of performance. Meta is playing a different game.

There is no single AI race

We talk about "the AI race" as though there is one finish line. There is not. There are several races happening at once, and the starting positions are almost the inverse of what you might expect.

OpenAI and Anthropic started with no platform and no distribution. What they had was the best models. OpenAI's ChatGPT became the most recognised AI product in the world through sheer product quality and timing. Anthropic's Claude is making serious progress in enterprise business, with strong performance in coding, reasoning, and agentic work. Businesses building internal tools, writing code, and running complex workflows increasingly choose Claude for that work.

Microsoft, Google, and Meta started from the opposite position. Each has a massive platform with hundreds of millions or billions of users. What they lacked was models that could match OpenAI or Anthropic on raw capability. Microsoft partnered with OpenAI to close that gap. Google built Gemini and embedded it across Android, Chrome, and Workspace. Meta spent $14.3 billion assembling a team to catch up.

The question now is which advantage proves more durable. Do the best models attract platforms, or do the biggest platforms eventually build good-enough models? Time will tell. Both sides are moving towards each other. OpenAI is building consumer features and distribution. Anthropic is expanding its enterprise platform. Microsoft, Google, and Meta are pouring resources into model quality.

For now, the races remain distinct. The best model for coding is not the same as the best model for consumer search, which is not the same as the model that reaches the most people through social media.

What Meta is actually betting on

The clearest signal in the announcement was not the model itself. It was the shopping features. Meta teased product recommendations embedded directly in the Meta AI chatbot, pointing users to things they can buy.

That is the commercial thesis. AI embedded in everyday tasks, inside platforms people already use, driving engagement and eventually transactions. Meta also highlighted features like estimating calories from a photo and visualising how a product would look in your home.

A new Contemplating Mode runs multiple agents simultaneously to handle complex reasoning tasks. Meta compared it to Google's Gemini Deep Think and OpenAI's GPT Pro. The example they gave was planning a family holiday, with one agent drafting an itinerary while another finds activities for children.

Why this matters for UK SMEs

If you are running a small or medium-sized business, you do not need to pick a winner in the AI race. You need to understand which race matters to you.

If your priority is internal efficiency, coding, or building AI into your operations, the enterprise-focused models from Anthropic and OpenAI are ahead. If your customers find you through social media, Meta embedding AI into Instagram and WhatsApp will change how they interact with your brand. Both things can be true at the same time.

The practical question is not "which model is best?" It is "where are my customers, and where are my team's biggest time sinks?" The answers usually point to different tools.

Three things to consider now.

Your customers' AI habits are forming. If 3.5 billion people start using AI through Meta's platforms, that shapes expectations for how they communicate with businesses. Pay attention to how AI features appear in WhatsApp Business and Instagram over the coming months.

Your internal AI stack is a separate decision. The model your team uses to draft proposals, analyse data, or build automations does not need to be the same one your customers interact with. Choose based on the task, not the brand.

Do not wait for the race to end. It will not. The companies building AI are competing on different fronts with different starting advantages. Your job is to identify which fronts affect your business and act on those.

Getting started

At gecco, we help UK SMEs work through exactly this question. Which AI tools fit your business, where to start, and how to build the skills and processes that make adoption stick. The technology matters, but the people and habits around it matter more.

If you want a clear picture of where your business stands, take the AI readiness assessment.