Meta teases 'Watermelon,' a compute-heavy Muse Spark update built for coding

Meta is preparing a new version of its flagship Muse Spark model, and this one is aimed squarely at people who write software. Chief AI Officer Alexandr Wang posted on X that the update — codenamed Watermelon — will bring "big improvements in coding and agentic capabilities to be more competitive with other leading models," and that it will arrive "soon" through Meta AI and a new API. No firm date was given.

What's actually new

Two things stand out. First, the update is expensive to run: Wang and follow-up reporting describe Watermelon as using far more compute than the original Muse Spark — one account puts it at an "order of magnitude" more computing capacity. That is Meta betting that scaling up training and inference compute translates directly into better coding.

Second, the focus on "agentic" work matters. An agentic model isn't just answering questions — it plans, runs tools, edits files, and checks its own output across many steps. That is exactly the capability that powers tools like Claude Code and OpenAI's Codex, and it's where Meta has trailed.

Where it lands against rivals

The original Muse Spark, which debuted in April 2026, scored 52.5% on SWE-Bench Pro, a benchmark that measures how often a model can resolve real GitHub issues. For context, Computerworld reports that GPT-5.5 reached 58.6% on the same test and Claude Opus 4.8 hit 69.2%. Citing anonymous sources, Business Insider reported that Watermelon has "already caught up with" GPT-5.5 — which, if it holds up on release, would close most of Muse Spark's gap with OpenAI while still leaving Anthropic's top model ahead.

Treat those numbers with care: the catch-up claim comes from unnamed sources, not a published benchmark, and Meta has not released Watermelon's own scores. Vendor benchmark claims routinely look different once independent testers run the same tasks.

Why it matters for builders

Meta is reportedly weighing an AI infrastructure and coding service that would compete directly with Claude Code and Codex. A capable, lower-cost model from a company that can afford to price aggressively would be good news for anyone learning to build with AI: more competition tends to mean cheaper tokens and more generous free tiers. Wang, asked when Meta would match Anthropic's Claude Opus 4.8, said only that it would happen "pretty soon."

The honest read for now: this is a signal, not a shipped product. There's no release date, no public benchmark, and no API you can call yet. If you're picking a coding assistant this month, choose on what's available and measured today — Claude, GPT, and the established agents — and revisit when Watermelon actually lands and independent testers weigh in. But the direction is clear: the biggest labs are now competing hardest on exactly the agentic coding workloads that vibe coders depend on.

Sources

• SiliconANGLE — Meta to release new AI model with advanced coding capabilities 'soon'
• Computerworld — Meta's AI chief says new Muse Spark update will sharpen coding, agentic AI
• Meta AI — Introducing Muse Spark