Meta Announces Major Coding and Agentic Upgrade for Muse Spark Model

Meta is preparing another update to its Muse Spark model, this time focused on two areas where the company has clearly lagged behind competitors: coding and autonomous agentic performance. The announcement came from Meta's head of AI, Alexandr Wang, who signaled that the upcoming version is meant to meaningfully narrow the gap with OpenAI and Anthropic.

Wang kept the announcement short and to the point, not mincing words about where the current model falls short.

Our next Muse Spark update is coming soon. Major improvements to coding and agentic capabilities, to better compete with other leading models - Alexandr Wang, Meta's head of AI

Where the Gap Comes From

The current version of Muse Spark, launched as Meta Superintelligence Labs' first product, is natively multimodal and can use tools as well as carry out visual chain-of-thought reasoning. Even so, Meta itself acknowledges in materials accompanying the launch that the model does not yet match rivals on long-horizon agentic tasks and coding work. A score of 59.0 points on Terminal-Bench 2.0, a benchmark that tests autonomous code generation and execution, compared to 68-75 points for competing models, shows the scale of the gap.

By comparison, the base model scores 58 percent on the Humanity's Last Exam test in its deeper-reasoning mode and 38 percent on FrontierScience Research, putting it roughly on par with Gemini Deep Think and GPT Pro in pure reasoning terms. The problem lies elsewhere: in multi-step tasks that require using a terminal, editing files, and maintaining context over extended periods, exactly where competition among coding assistants is decided today.

Watermelon as the Answer

The Watermelon codename refers to the version meant to close that gap. Unofficial information cited by Business Insider indicates that the new model uses substantially more compute than its predecessor and, in internal testing, achieved results close to OpenAI's GPT-5.5. Meta has not officially confirmed these figures, but the fact that internal comparisons leaked to the press at all suggests the company wants to manage expectations ahead of launch.

This is already the second major Muse Spark update in a short span, illustrating the pace at which Meta is trying to make up ground after its language models spent most of 2025 in the shadow of Gemini, Claude and GPT. At the same time, the company has moved away from the fully open distribution model that characterized the Llama series, which some commentators read as a sign that Meta now takes the race for agentic capability more seriously than open-source code.

What It Means for Businesses

Analyst Pareekh Jain notes that even partially closing the gap could shift the balance of power in the enterprise market. Greater competition in the coding and agentic model segment typically means lower prices and less vendor lock-in for companies, and Meta, by offering the model free in its consumer version while preparing an API, is targeting exactly those enterprises that today must choose between OpenAI and Anthropic.

Technical success, however, is only half the equation. According to analysts cited by InfoWorld, for Muse Spark to genuinely threaten Cursor, Claude Code or Codex in production environments, Meta needs to demonstrate not just benchmark results but also reliable execution of agentic tasks, robust safeguards, and a surrounding tool ecosystem. The absence of a public API at launch and the restriction of access to select partners suggest the company itself views this stage as a test phase rather than a finished product for developers.

For Polish development teams already testing Claude Code, Cursor or GitHub Copilot, the upcoming Muse Spark update is for now a signal to watch rather than a reason to switch tools. Until the model reaches a broader API with clear pricing, it is hard to judge whether the touted jump in Terminal-Bench will translate into real quality gains on actual production code, rather than just a better synthetic benchmark score.

Meta has not yet given a specific release date for Watermelon or pricing for the API. Given the pace of announcements in recent months, the update could arrive within the coming weeks, though the company has a track record of pushing back timelines when internal safety and performance tests fall short of targets.

Sources: Meta AI Blog (ai.meta.com), InfoWorld (infoworld.com)