Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Meta has launched its first image-generation model since chief executive Mark Zuckerberg undertook a dramatic overhaul of the company’s AI efforts, as it races to catch up with rivals including OpenAI and Google.
The $1.6tn social media platform on Tuesday revealed Muse Spark Image, which it said used “advanced reasoning to understand complex prompts” and would be integrated into its own Meta AI chatbot.
The model would be used to power new editing features on Meta’s Instagram photo app and integrated into the suite of tools used by marketers to generate advertising for the platform, the company said.
The launch comes three months after Meta launched its first Muse Spark model, which was multimodal — meaning it can process various types of data — but produced text-only outputs.
Meta also said on Tuesday that it had a new video model in development called Muse Video.
Over the past year, Zuckerberg has poured billions of dollars into developing AI infrastructure and poaching top AI talent in an effort to catch up with rivals Google, OpenAI and Anthropic.
He tasked a new team of AI researchers in a secretive lab called “TBD” — To Be Determined — to focus on spearheading the development of state of the art models. It is led by Scale AI founder Alexandr Wang.
The push came after its Llama 4 large language model, released in April 2025, lagged rivals. Meta also struggled to develop cutting-edge image and video models. It instead opted last year to strike a multiyear, multibillion-dollar deal paying the AI start-up Midjourney to use its image-generation technology.
Meta will now wind down its use of Midjourney for certain image-generation capabilities, replacing it with Muse Spark Image, according to one person familiar with the matter. Midjourney will continue to power some of Meta’s video generation features, but this would also be replaced by Muse Video over time, the person said.
Zuckerberg has faced increasing pressure to justify the company’s ballooning AI spending and recently slashed 10 per cent of its workforce, hurting staff morale.
Meta shares jumped almost 9 per cent in a single session last week on the news that the company is exploring setting up a cloud-computing business in order to sell off any surplus computing power.
The original Muse Spark is widely considered an improvement on Llama 4, but Zuckerberg and Wang acknowledged that it was still behind on coding and agentic capabilities that power AI bots to operate independently.
Unlike Meta’s previous Llama models, which have been “open” — or freely available for developers to tweak — Muse Spark is a smaller, closed model, making it harder to assess. Meta promised some businesses and developers would have access to the model via a private application programming interface, but this has been repeatedly delayed.
Muse Spark Image beat leading models from Google and Elon Musk’s SpaceXAI in benchmarks for text-to-image and editing, according to rankings on Arena. Arena compiles leader boards from votes by testers who have early access on the independent platform, who ranked Muse Spark Image below OpenAI’s GPT Image 2 model, released in April.
Additional reporting by Cristina Criddle in San Francisco
Credit: Source link
