Meta Teases Coding-Focused AI Model Update as Watermelon Trains on 500,000-GPU Cluster

Quick Facts

• Meta Chief AI Officer Alexandr Wang said a Muse Spark update with improved coding and agentic capabilities will arrive “pretty soon.”
• Meta’s next-generation model, codenamed Watermelon, is currently in training on the Prometheus cluster in New Albany, Ohio, using roughly 500,000 GPUs.
• Meta projects $125 billion to $145 billion in capital expenditure this year, up from an earlier forecast of $115 billion to $135 billion.

Meta Platforms is preparing to ship an updated version of its Muse Spark AI model with meaningfully stronger coding capabilities, according to Chief AI Officer Alexandr Wang. Wang made the announcement on X on July 3, responding to a user who asked when Meta would release a model competitive with Anthropic’s Claude Opus 4.8 on coding tasks.

Wang said it would happen “pretty soon” and added that users would like what the company had “cooking.” He also confirmed that the upcoming Muse Spark update is significantly more capable at coding than the current version.

Where Meta Stands on Coding Benchmarks

The original Muse Spark scored 52.5% on SWE-Bench Pro, a benchmark that tests AI systems on coding tasks drawn from actively maintained repositories. GPT-5.5, OpenAI’s current flagship, reached 58.6% on the same test. Business Insider reported that Meta’s new algorithm has reached parity with GPT-5.5 across several closely watched benchmarks.

Anthropic’s Claude Opus 4.8, released May 28, 2026, scored 69.2% on SWE-Bench Pro, which is 10.2 percentage points higher than GPT-5.5. Claude Opus 4.8 trails GPT-5.5 on Terminal-Bench 2.0, a separate coding evaluation.

Watermelon: The Bigger Play

Beyond the near-term Muse Spark update, Wang told employees at an internal town hall on July 2 that a far larger model is already in training. The model, codenamed Watermelon, is the successor to Muse Spark, which Wang called an “appetizer.”

Wang said Watermelon uses an order of magnitude more compute than Muse Spark. The model runs on Meta’s Prometheus cluster, a facility under construction in New Albany, Ohio, drawing over a gigawatt of power and housing an estimated 500,000 GPUs. Meta has not set a public release date.

The compute jump comes with a cost. Meta told investors this year it expects to spend $125 billion to $145 billion on chips, data centers, and related infrastructure, raising its earlier forecast of $115 billion to $135 billion.

Zuckerberg’s Candor

CEO Mark Zuckerberg acknowledged at the same town hall that Meta’s AI investments have not paid off as quickly as expected. He told employees the company should see more substantial returns within three to six months, pushing the timeline toward late 2026.

Zuckerberg also conceded he made mistakes during a recent workforce restructuring and said he would “almost certainly make more.” The comments reflect pressure on Meta’s Superintelligence Labs unit, which was created more than a year ago.

A Closed-Source Shift and Cloud Ambitions

Muse Spark marked a departure from Meta’s longstanding open-source model strategy under the Llama brand. Truist analysts noted in April that the closed-source approach signals a push toward high-performance, specialized infrastructure rather than open community development.

Meta is also developing plans to sell access to its AI computing power and models as a cloud service, according to Bloomberg. Gil Luria, managing director at D.A. Davidson, said the move would pressure neocloud providers more than major hyperscalers. “Those companies like CoreWeave and Nebius rely on Meta for their growth and Meta may not need them anymore,” Luria said.

One caveat: Wang’s benchmark parity claim for the new model came during a private internal meeting. No independent evaluation or published model card has been released. The assertion cannot be confirmed externally until Meta submits the model to third-party testing.

Read more: Meta to release new AI model with advanced coding capabilities ‘soon’