What Is Meta Watermelon? 10× More Compute

10× More Compute: A First Look at Meta Watermelon

Meta Watermelon is an unreleased large language model codename associated with Meta Superintelligence Labs, reportedly intended to follow Muse Spark and trained with roughly 10× more compute. The codename entered public discussion through reports about a July 2026 internal town hall, and a September 2, 2026 post said Mark Zuckerberg had teased the model in a public social-media post. Meta has not publicly released Watermelon, its weights, a model card, a benchmark table, a price, an API, or a confirmed launch date.

Meta Watermelon is currently a codename for an unreleased Meta frontier model, not a publicly available product.

The main reported claim comes from Alexandr Wang, Meta’s chief AI officer, who told employees that Watermelon had caught up with OpenAI’s GPT-5.5 on closely followed benchmarks. The benchmarks were not named, the scores were not published, and independent evaluators have not tested the model. The most defensible description is a large pretraining project with an ambitious internal performance claim.

Key Takeaways

• Meta Watermelon is reportedly the next model after Muse Spark, whose internal codename was Avocado.
• The model is still in training according to the available reporting, so its final architecture and behavior may change.
• Watermelon reportedly uses an order of magnitude more compute than Muse Spark, meaning roughly 10× under ordinary numerical usage.
• Meta leadership reportedly claimed parity with GPT-5.5 on unnamed internal benchmarks.
• No public benchmark results, context window, parameter count, price, API, weights, license, or release date has been confirmed.
• Community timing posts range from “not released” in August 2026 to a possible second-half-of-September 2026 window.

What Is Meta Watermelon?

Meta Watermelon is the reported codename for Meta’s next flagship model after Muse Spark. The model is associated with Meta Superintelligence Labs, the division responsible for Meta’s frontier AI work. Available reports describe it as a large language model rather than an image, video, speech, or specialized narrow model, but Meta has not published an official modality specification.

The model’s predecessor is important. Muse Spark launched in April 2026 under the internal codename Avocado. It performed well on some evaluations, but third-party reporting placed it behind the leading OpenAI, Anthropic, and Google systems overall. Watermelon therefore represents Meta’s attempt to move from a strong model in selected tests toward a broadly competitive frontier system.

A July 2026 report cited two people familiar with an internal town hall. According to that account, Wang said Watermelon was still in training, used roughly 10× more compute than Avocado, and had caught up with GPT-5.5 on certain benchmarks. The report did not identify the tests or provide numerical results. Meta did not issue a public technical announcement in the supplied evidence.

The available evidence describes Meta Watermelon as a training project and internal milestone, not a released model family.

The scale claim also fits Meta’s broader infrastructure strategy. Reports described Meta’s 2026 AI capital expenditure guidance as between $125 billion and $145 billion. A third-party account connected the training effort with the Prometheus computing facility, described as a 1-gigawatt site with an estimated 500,000 GPUs. Those infrastructure figures provide context for Meta’s compute strategy, but they do not prove Watermelon’s exact training budget.

For background on the products Meta has actually shipped, see Meta’s Muse Image and Muse Video: What MSL’s First Generative Models Actually Ship.

Meta Watermelon at a Glance

Specification	Current evidence
Developer	Meta Superintelligence Labs
Codename	Watermelon
Reported predecessor	Avocado, the internal codename associated with Muse Spark
Type	Large language model in training, based on reporting
Modality	Not yet confirmed
Context window	Not yet confirmed
Reported compute	Roughly 10× more than Muse Spark, unconfirmed
Reported benchmark position	Comparable to GPT-5.5 on unnamed internal benchmarks, unverified
Pricing	Not yet confirmed
Availability	In training and not publicly released in the supplied evidence
API access	Not yet confirmed
Model weights	Not yet confirmed
License	Not yet confirmed
Parameter count	Not yet confirmed
Release date	Not yet confirmed

How Meta Watermelon Works / What Makes It Different

The available information says more about Watermelon’s scale than its design. There is no public architecture description, parameter count, training-data summary, inference method, reasoning mode, or tool-use specification. Any claim about mixture-of-experts routing, multimodal encoders, long-context attention, or a named reasoning system would go beyond the evidence.

The clearest analytical label is Order-of-Magnitude Compute. Wang reportedly used that phrase to compare Watermelon with Avocado. In ordinary numerical language, an order of magnitude means a factor of approximately 10, but the comparison method is unknown. It could refer to total training operations, accelerator hours, hardware count, training duration, or a combination of measures.

A reported 10× increase in training compute describes resource scale, not a guaranteed 10× increase in capability.

Compute scaling can improve model quality, but gains are not linear. More hardware does not automatically produce better data quality, better post-training, safer behavior, or more reliable tool use. The final result also depends on data mixture, optimization, evaluation design, reinforcement learning, and deployment constraints.

The second anchor is Internal Benchmark Parity. Meta’s reported claim concerns parity with GPT-5.5 on “closely followed” benchmarks, but no benchmark names or scores are available. A model can match a rival on selected mathematics or coding tests while performing differently in factuality, agent reliability, multimodal understanding, latency, or cost.

The third is Training-Stage Uncertainty. Watermelon was described as still training. Intermediate checkpoints can change substantially before release, especially after instruction tuning, safety work, tool integration, and product optimization. A result observed during training cannot be assumed to describe the eventual public model.

The fourth is Agentic Capability Target. Reports surrounding Meta’s roadmap pointed to stronger coding and AI-agent performance as important goals. That indicates where Meta wants to compete, not what Watermelon can currently do. No public hands-on demonstration establishes autonomous task completion, computer use, coding reliability, or multi-step planning.

What You Can Do With Meta Watermelon

At present, the practical answer is that developers cannot reliably use Watermelon. No official access path has been identified, and no public weights or developer documentation are available.

If Meta eventually releases it, the reported target areas suggest several likely use cases:

• Software development: Wang reportedly connected upcoming Meta model improvements with stronger coding capability. Watermelon may therefore be evaluated for code generation, debugging, repository navigation, and software-agent tasks.
• AI agents: Reports described improved agent performance as a future objective. Relevant tests would include planning, tool calling, browser or computer control, and long-horizon task completion.
• General-purpose assistance: As a reported flagship large language model, Watermelon could eventually support writing, analysis, question answering, and enterprise workflows.
• Meta product integration: A frontier model could be applied across Meta’s consumer products, but no Watermelon-powered Facebook, Instagram, or WhatsApp feature has been confirmed.

These remain potential uses, not demonstrated capabilities. Teams should not select Watermelon for production based on the 10× compute claim or the unnamed benchmark comparison alone.

How Meta Watermelon Compares

The most useful comparison is currently about evidence and availability rather than raw capability.

Model	Evidence status	Public access position	Relevant comparison
Meta Watermelon	Internal parity claim, independently unverified	Not yet confirmed	Reportedly uses roughly 10× Muse Spark’s compute
GPT-5.5	Publicly released by April 2026, according to reporting	Available through OpenAI products and API, according to reporting	Named as the comparison target
GPT-5.6	Limited preview reportedly began on June 26, 2026	Access was described as restricted in supplied reports	Shows that the frontier target moves during Watermelon’s training
Muse Spark	Released on April 8, 2026	Public predecessor model	Received a reported 52-point Intelligence Index score from an independent benchmarking firm

The reported GPT-5.5 parity remains an internal claim until Meta publishes tests that others can reproduce.

This makes comparisons with GPT-5.6 especially provisional. The supplied evidence says Watermelon reportedly matched GPT-5.5, while GPT-5.6 had already moved the frontier into a limited preview. A model can catch an earlier target and still trail the strongest available system.

For teams that need an accessible comparison point while waiting for official Watermelon information, GPT-5.5 is listed as a chat model with public product and API availability described in the supplied reporting.

Availability: How to Access Meta Watermelon

Meta Watermelon is not confirmed to be accessible through Meta AI, an official API, a research preview, or downloadable weights. The supplied evidence contains no official model identifier, endpoint, SDK, waitlist, or developer documentation.

Community timing remains unsettled. On September 1, 2026, one tracker marked Watermelon as not released in August, while another roundup placed it in a possible second-half-of-September window. Those posts are not launch notices. They are useful only as indicators of community expectations, as shown by the release-tracking post and the possible timing roundup.

The correct access workflow is to wait for a Meta announcement, official documentation, or a model card. A genuine release should establish the model name, supported modalities, usage limits, safety terms, evaluation results, and pricing. Until then, third-party claims that Watermelon is “ready” do not provide a usable access route.

What We Don’t Know Yet

Several questions remain open:

• What architecture does Watermelon use?
• How many parameters, tokens, accelerators, or training operations are involved?
• Does “10× compute” refer to the full training run or one part of it?
• Which benchmarks supposedly matched GPT-5.5?
• Were the tests run on a final model, an intermediate checkpoint, or a special internal configuration?
• What are Watermelon’s context window, modalities, latency, and tool-use capabilities?
• Will Meta release weights, an API, a hosted assistant, or only selected product integrations?
• Is the model open source, open weights, or closed?
• What price and rate limits will apply?
• When will independent evaluators and developers gain access?

A September 1, 2026 community post predicted that Watermelon could be “Fable-level” and open source, but that prediction has no disclosed test or official support. It should not be treated as a specification.

Frequently Asked Questions

Is Meta Watermelon open source?

Meta Watermelon is not confirmed to be open source. No weights, license, or official distribution terms have been published in the available evidence.

How much does Meta Watermelon cost?

Meta Watermelon has no confirmed price. The model is still described as being in training, with no public API or product pricing available.

When will Meta Watermelon be released?

Meta Watermelon has no confirmed release date. Community trackers have suggested possible timing in the second half of September 2026, but that remains unconfirmed.

What is Meta Watermelon’s context window?

Meta Watermelon’s context window has not been confirmed. No public model card or technical specification lists its token capacity.

Meta Watermelon vs GPT-5.5?

Meta Watermelon is reportedly comparable to GPT-5.5 on unnamed internal benchmarks, but no public scores or independent tests establish parity. GPT-5.5 has the stronger public evidence package because its model access and evaluation results are available.

Is Meta Watermelon available through an API?

Meta Watermelon is not confirmed to be available through an API. The supplied evidence identifies no official endpoint, developer documentation, or public access channel.

What to watch next

The strongest update signals are an official Meta model announcement, a named benchmark table with evaluation conditions, and a real access mechanism such as an API or downloadable weights. Pricing, context length, license terms, and independent testing will determine whether Watermelon is a frontier product or only an internal milestone.

Building similar frontier chat workflows? On kie.ai you can try GPT 6.1 Sol, Claude Sonnet 5.5, and Claude Opus 4.6.