/
Navigation
Chronicles
Browse all articles
Explore
Semantic exploration
Research
Entity momentum
Nexus
Correlations & relationships
Story Arc
Topic evolution
Drift Map
Semantic trajectory animation
Posts
Analysis & commentary
Pulse API
Tech news intelligence API
Browse
Entities
Companies, people, products, technologies
Domains
Browse by publication source
Handles
Browse by social media handle
Detection
Concept Search
Semantic similarity search
High Impact Stories
Top coverage by position
Sentiment Analysis
Positive/negative coverage
Anomaly Detection
Unusual coverage patterns
Analysis
Rivalry Report
Compare two entities head-to-head
Semantic Pivots
Narrative discontinuities
Crisis Response
Event recovery patterns
Connected
Search: /
Command: ⌘K
Embeddings: large
TEXXR

Chronicles

The story behind the story

days · browse · Enter similar · o open

Sources: Alexandr Wang said Meta's model currently in training, codenamed Watermelon, matches GPT-5.5 and uses an “order of magnitude more compute than Avocado”

Business Insider

Context & Ripple Effects

Meta’s model effort has moved quickly from Avocado—described internally as its most capable pre-trained base model and more compute-efficient than Llama 4 Maverick—to Watermelon, a larger model still in training. Earlier coverage also indicated Meta was preparing Wang-era models for open-source release.

The reported jump matters because it pairs a claimed frontier-level capability target with a stated order-of-magnitude increase in training compute, while Meta is also arranging substantial data-center financing and incorporating Manus for agent delivery across its products.

First-order effects

  • Meta’s AI organization faces materially higher infrastructure and capital demands for Watermelon than for Avocado, assuming the reported compute increase holds.
  • A claimed match with GPT-5.5 would strengthen Meta’s near-term positioning in frontier-model competition, but the comparison remains a source-reported internal assessment rather than an independently established benchmark.

Second-order effects

  • Meta’s data-center buildout and financing become more directly tied to model competitiveness: greater training-compute requirements raise the strategic value of secured capacity and make execution on projects such as Hyperion more consequential.
  • If Meta pairs stronger base models with Manus-derived agents and distribution through Meta AI, product teams can shift from model availability toward deploying more capable agent features across Meta’s services.

Third-order effects

  • The progression from Avocado’s efficiency emphasis to Watermelon’s much larger compute budget illustrates that efficiency gains are not eliminating the race for large-scale training infrastructure; they may instead enable companies to spend more aggressively at the frontier.
  • Meta’s reported intent to offer some Wang-era models under an open-source license could make the eventual release strategy consequential: a strong openly available model would widen access to capable systems, while a closed release would reinforce the concentration of frontier capability among infrastructure-rich firms.

The trend: Frontier AI competition is increasingly being decided by the combination of model efficiency, access to financed compute infrastructure, and the ability to turn base models into widely distributed agent products.

Discussion

  • @alexandr_wang Alexandr Wang on x
    First, Mark was clearly talking about the industry's progress on agentic capabilities on the whole.  But, while we're on the topic: Our next Muse Spark update is coming soon.  Big improvements in coding and agentic capabilities to be more competitive with other leading models.  E…
  • @kimmonismus @kimmonismus on x
    Meta is expected to release an Opus model in the very near future.  Reportedly, Mark Zuckerberg's disappointment with Agentic AI was directed at the field of agentic AI in general rather than at Meta specifically.  What exactly he meant by that remains unclear.  Personally, I don…
  • @lichangchen2 Lichang Chen on x
    I cannot wait for the debut of our coding agents with computer use on public harnesses like OpenCode and Cursor. I've been testing it for a while. Go Muse Spark!
  • @cryptopunk7213 @cryptopunk7213 on x
    wow, so both meta and spaceX have 10-15T models currently being trained. people forget ceos like zuck and elon have deep pockets that will scale data center compute quicker than anyone else if we assume compute = better model then there's a timeline where meta and spaceX catch
  • @bindureddy Bindu Reddy on x
    Meta says they have a new model in training, that's as good as GPT 5.5 The problem - the world has already moved on to GPT 5.6 Plus unlike Grok or Gemini, no one can even verify their model performance🤷‍♀️
  • @charlesrollet1 Charles Rollet on x
    SCOOP: Alexandr Wang says Meta's upcoming AI model - codenamed Watermelon - has caught up to OpenAI's GPT-5.5. “Watermelon uses an order of magnitude more compute than Avocado,” he said in an internal meeting, referencing Meta's previous AI model. https://www.businessinsider.com/…
  • @kimmonismus @kimmonismus on x
    According to Alexandr Wang, Meta's next frontier model, codenamed Watermelon, has already caught up with OpenAI's GPT-5.5 on internal benchmarks. The model is still training. One interesting detail: Watermelon reportedly uses an order of magnitude more compute than Avocado [image…
  • @gergelyorosz Gergely Orosz on x
    The industry's progress on agentic capabilities is pretty incredible, esp the last few months. Cloud coding agents are starting to become more common and capable, and local ones are already v powerful. What do they not see at Meta that eg I do?
  • @charlesrollet1 Charles Rollet on x
    During the same meeting, Mark Zuckerberg said that Meta is on a “journey to superintelligence.” He added that generally, progress on AI agents technology has been slower than expected. https://www.businessinsider.com/ ...
  • @firstadopter Tae Kim on x
    Thank you @BusinessInsider All you FUD fearmongers on Meta's capex can calm down now “Watermelon uses an order of magnitude more compute than Avocado” Same town hall everyone freaked out over from sensationalized headlines. Clown show. [image]
  • @rwang07 Ray Wang on x
    🤣
  • @rvc330 @rvc330 on x
    when you're a frontier lab but also selling compute because you don't need it but also reassigning researchers to data labeling and also looking into prediction markets and VR glasses and the metaverse
  • @dratchcap Matt Dratch on x
    The more I use harnesses and agents, the more I think web pages (and apps) are the Yellow Pages of our time.  You can feel a first-principles redesign of the internet, GUIs, and operating systems coming.  In that world, $META has a real shot to gain eyeball time and become the ne…
  • @inductionheads @inductionheads on x
    It isn't just that wang is a huckster It's that he literally led them down exactly the wrong path (all this screen recording nonsense) to try to retcon his scale AI thesis into relevance Time to cut bait Mark (hire @willccbb instead)
  • @scaling01 @scaling01 on x
    Meta is still training their big models but we will get an update on Muse Spark soon
  • @andrewcurran_ Andrew Curran on x
    Alexandr Wang, head of the META Superintelligence Lab, says that a new Muse Spark update, and an Opus level Muse variant, are both on the way. [image]
  • @pythiar @pythiar on x
    In other words, they're not catching up fast enough for Zuck but they are catching up.
  • @alexeheath Alex Heath on x
    Watermelon uses an order of magnitude more compute than Avocado
  • @daniellefong Danielle Fong on x
    Just wait until you see the amount of compute used for the great pumpkin
  • @edzitron Ed Zitron on x
    Hey everyone we made a model potentially as good as someone else's. It also takes up way more compute. No real clue how we'll monetize it, and we haven't been able to stand up an API service for our current one. Anyway, see ya
  • Dylan Jones Dylan Jones on linkedin
    Earlier this year, Meta announced that they were laying off thousands of employees to make room for a huge investment in AI. …
  • @sungkim Sung Kim on bluesky
    In U.S., AI race is dominated by 2.5 players (.5 being Google).  Will we be returning to 3.5 players soon?  —  www.businessinsider.com/meta-ai- mode...
  • @petertl Peter Thal Larsen on bluesky
    Mark Zuckerberg in January: “We're starting to see projects that used to require big teams now be accomplished by a single very talented person.”  —  Mark Zuckerberg in July:  —  www.reuters.com/business/zuc...