Watermelon: A New Generation Using Ten Times More Power

Alexandr Wang, who leads research within Meta's Meta Superintelligence Labs division, wrote on X on Thursday that the next update to the Muse Spark model is on its way and will bring "major improvements in coding and agent capabilities to be more competitive against other leading models." He was responding to earlier statements by Mark Zuckerberg, who admitted at a company meeting that the development of AI agents was progressing slower than Meta had expected.

According to Business Insider, which cited two anonymous sources directly from the meeting, Wang told employees at the same meeting that Watermelon — the internal designation for the next version of Muse Spark — "uses an order of magnitude more computational power than Avocado" (which was the internal name for the first Muse Spark released in April 2026). And more importantly: according to Wang, Watermelon has already caught up with OpenAI's flagship model GPT-5.5 in benchmarks.

One key detail: Wang did not specify which particular benchmarks he was referring to. This is significant because different tests measure different capabilities — from mathematical reasoning to code writing to general knowledge. Without specification, it remains an open question what exactly Watermelon excels at.

Why Coding and Agents Matter More Than Benchmarks

While the general public is mainly interested in how well AI answers questions, two things are crucial for businesses and developers: the ability to reliably write and fix code and the ability to independently perform multi-step tasks (i.e., agent behavior). It is in these disciplines that models from OpenAI and Anthropic have so far reigned supreme — especially Claude Opus is considered the gold standard for programming among developers.

When Wang was asked on X when Meta would have a model on par with Claude Opus for coding, he replied that "quite soon" and that we have something to look forward to. If Watermelon can truly compete with Claude Opus in real-world programming, it would be a pivotal moment — Meta would move from being an "open-source model for experiments" to a full-fledged tool for professional development.

What This Means for Businesses and Czech Developers

Pareekh Jain, an analyst at Pareekh Consulting, points out that a strong model from Meta could increase competition and lower the prices of AI tools for businesses. "If Watermelon is offered as an open-weight model or at a low price, it could make AI assistants for programming accessible to smaller companies while also improving data control and reducing dependence on a single vendor," he stated.

This is also relevant for the Czech market. Czech companies and developers currently most often reach for Microsoft's Copilot (based on OpenAI) or Anthropic's Claude. Both tools typically cost $10–30 per month per developer, with prices multiplying for team licenses. A Meta model, if available for free or significantly cheaper — for example, via Meta AI, which is currently free — could significantly shift the balance, especially for smaller Czech companies and startups.

But beware: Meta AI is not yet officially available in Czech and faces regulatory hurdles in the EU. The service is available to a limited extent in Europe, which could hinder wider adoption. On the other hand — if Meta releases the model as open-weight (as it did with the Llama models), Czech developers could use it locally without restrictions.

Betting on Cloud and Agent Platform

According to CIO.com, Meta is preparing its own cloud infrastructure through which it would offer computational power and access to its models to companies. Meta's total investments in chips, data centers, and infrastructure will reach $125 to $145 billion this year — more than some countries spend on their entire military.

Last year's acquisition of the startup Manus for approximately $2 billion also fits into the mosaic. Manus developed AI agents that could handle everything from job interviews to stock analysis. Charlie Dai, an analyst at Forrester, sees a clear signal in these steps: "Meta seems to want to go beyond basic models and become a platform for creating AI-native applications and agents."

Obstacles on the Path to Businesses

However, analysts' enthusiasm is tempered by the reality of corporate IT. For Meta to succeed among enterprise customers, according to Charlie Dai, it must demonstrate reliable code quality in real-world conditions, secure execution of agent tasks, and strong governance and management. In addition, it needs to build a vibrant ecosystem of developers around its models.

For European and Czech companies, there is also a double burden: in addition to technical quality, they must also address compliance with the European AI Act regulation. Models from American companies undergo stricter assessment in the EU, and companies that deploy them are responsible for their outputs. This favors models that can be operated on their own infrastructure — and Meta's open-weight models offer this option.

According to Wang, the new model will be available "soon" via Meta AI and a new API. A specific date has not yet been announced.

Muse Spark vs. Competition: Market Comparison

The current market for AI models for coding looks as follows:

• OpenAI GPT-5.5 / Codex: Widely used, integrated into GitHub Copilot and ChatGPT. Price: from $20/month (ChatGPT Plus) to $200/month (Pro).
• Anthropic Claude Opus 4.7: Considered best for complex coding. Price: from $20/month (Claude Pro).
• Google Gemini 2.5: Strong in mathematics and logic, integrated into the Google ecosystem. Price: from $20/month (Google One AI Premium).
• Meta Muse Spark / Llama 4: Free open-weight models, but previously lagged in coding. Watermelon should change that.

If Watermelon truly catches up with the top and remains available as open-weight or at a low price, it could become the most interesting alternative for developers who do not want to pay for subscriptions — including Czech programmers, freelancers, and small teams.