Meta Watermelon After Meta Connect: The Keynote Came and Went Without It

Meta Watermelon was supposed to be the reason to watch Meta Connect 2026. The keynote ran on September 23 at Menlo Park, and the model — a working codename for Meta's next frontier-scale system, teased by Mark Zuckerberg on September 2 with a single watermelon emoji — was not in it. No new frontier model was announced at all. Neither did the other half of that September 2 promise land: the open weights for Muse Spark 1.2 are still not on Hugging Face, now forty-five days after Zuckerberg said they were coming "in the coming weeks." What Connect did deliver was Muse product surface rather than a new brain — a Muse Realtime Avatar model for live video conversation, a customizable Muse Voice, and a Connectors platform that opened to developers on September 19 — while Muse Spark 1.3, Meta's current frontier checkpoint, remains the newest frontier model Meta actually ships.

This is still a what-we-know-so-far piece, but its first checkpoint has now resolved. The September 2 tease set up two expectations and a venue; the venue has passed and both expectations are still open. Everything below about Watermelon's capability, timing and distribution remains either vendor- or exec-stated, or drawn from press reports of internal Meta meetings that Meta has never confirmed. What changed this week is not the model. It is the calendar: the most likely window for a public reveal has closed without one, which turns the reported October target from a footnote into the next dated thing to watch.

What the tease actually said

The sign-off line, relayed on X by the account @scaling01, reads: "Next up 🍉 and Muse Spark open weights releases coming soon." Read literally, it names two upcoming things: a model whose public placeholder is a watermelon emoji, and the open-weight release of the Muse Spark line. The context matters. The post it ended was a launch — Muse Spark 1.3, Meta's fourth Muse Spark checkpoint of 2026, priced unchanged at $1.25 per million input tokens and $4.25 per million output tokens, with Meta claiming roughly 25% fewer tokens per task than Muse Spark 1.2. All of those figures are either Meta's own or Artificial Analysis's; none of that is what made the tease newsworthy.

What made it newsworthy is that a CEO appended an unreleased flagship and an unfulfilled open-source promise to a launch post — an implicit roadmap, in public, with no dates. In the run-up to the keynote, the developer-watcher account @testingcatalog published a pre-event expectation list naming a new Watermelon-class model, Muse Spark open-sourcing, a Muse Video launch with API availability, and new voice models and voice-creation features as the things to expect. That list was an expectation, not a schedule, and the keynote did not deliver on its model half. A tease like this is not a spec sheet, but it remains the strongest public signal that Watermelon is real and next.

Meta Watermelon: the model behind the emoji

Meta Watermelon is a working codename, not a product name Meta has confirmed. What is known comes from four layers of reporting, none of them an official model card.

First, The Information reported on August 25, 2026 that Meta is targeting an October 2026 release for Watermelon, citing internal documents. The same reporting tied Watermelon to a broader consumer push — a companion AI agent platform reportedly codenamed Hatch, with a premium subscription reported at up to $199.99 a month — and noted that the internal documents did not clarify whether Watermelon would launch inside the Muse family, as a successor to Muse Spark, or as a standalone system. Second, a July 2, 2026 all-hands, first reported by Business Insider, has Alexandr Wang — Meta's chief AI officer and the head of Meta Superintelligence Labs — telling employees that Watermelon had "caught up" to OpenAI's GPT-5.5 on key benchmarks, using what he described as an order of magnitude more compute than the previous generation. The benchmarks were not named, Meta declined to comment, and no independent evaluation exists. Third, CNBC has since reported that Meta paused a pre-training stage in July and later resumed it, pushing the launch to no earlier than October. Fourth, on September 23, as the keynote ran, coverage relayed Wang saying Meta would soon release its most capable model yet, with the shipped Muse Spark line framed as an "appetizer" — exec-stated and reported secondhand, though the outlet did not date the remarks, and in any case not a product announcement.

• Working codename — Meta Watermelon; no confirmed product name, parameters, or architecture.

• Reported target — October 2026, per The Information's August 25 report citing internal documents; CNBC's pre-training-pause reporting points the same way.

• Compute — "an order of magnitude" more than the previous generation, per Wang's July 2 all-hands as reported; not independently verified. The training cluster is reported to be Prometheus, Meta's roughly 1-gigawatt New Albany, Ohio site, at around 500,000 GPUs.

• Capability claim — reportedly matches GPT-5.5 on unnamed internal benchmarks; no external run exists.

• Lineage — unclear: Muse-family successor, Llama-family release, or standalone; internal documents reportedly do not say.

• Distribution — unknown, including whether it will ship open-weight. Nothing at Connect changed that.

The "order of magnitude" compute figure deserves a skeptical note. Ten times the training compute of an already-large model does not imply a ten-times-better model — scaling laws are sublinear, and a tenfold compute increase typically buys a fraction of the capability gain. Treat the GPT-5.5-parity claim as a directional statement about Meta's compute posture, not a verified result.

What Meta Connect actually delivered

Meta Connect 2026 ran September 23–24 at Menlo Park. The keynote was about Muse, Meta's personal agent, and about hardware: camera-free Ray-Ban Meta Audio glasses at $349, a third-generation Ray-Ban Meta from $449, roughly 100-gram Meta VR Glasses at $1,299 for spring 2027, and the palm-sized Muse Charm device. On the model side, what shipped was narrower than the pre-event expectations.

• Announced at Connect — Muse Realtime Avatar, a model that drives an avatar's face, body and voice during a live video conversation. Meta described it as coming "in the next months," not as generally available.

• Announced at Connect — Muse Voice, a customization layer that lets a user change the assistant's voice, accent and speaking speed by describing what they want.

• Opened before Connect — Muse Connectors went to developers on September 19. Meta says it received more than 1,500 applications in under a week, with GitHub, Notion and Granola among the named partners.

• Not shipped — Muse Video. Previewed on July 7, 2026 alongside Muse Image and promised "soon to creators and in Meta AI," it still has no release date, no public API, no pricing, no resolution and no clip length. The only developer-access signal remains an unlaunched "Video Playground" section spotted in Meta's developer platform, which contains no endpoint, schema or model identifier.

• Not shipped — a voice-generation model or API. The "Voice Playground" trace is still unreleased, and the voice model Meta has actually shipped, Muse Voice Transcribe, is a transcription model rather than a voice-creation one.

• Not shipped — any new frontier model, and any Watermelon-class reveal.

The distinction most likely to be lost in the coverage is between Muse Voice and Muse Voice Transcribe: the first customizes how Meta's assistant sounds, the second turns speech into text. Neither is a voice-cloning API, and nothing at Connect made one available.

The open-weights half is now six weeks old

The second clause — "Muse Spark open weights releases coming soon" — is not new. On August 10, 2026, Zuckerberg announced that Meta would release the weights of Muse Spark 1.2 "in the coming weeks," a reversal of the closed, invitation-only strategy Meta had taken with Muse Spark's April debut. The same day, Meta released Muse Glimmer, a 30-billion-parameter open-weight model under an Apache 2.0 license, distilled from Muse Spark's outputs and aimed at a single consumer GPU or laptop-class hardware envelope. Muse Glimmer shipped; Muse Spark 1.2's weights did not.

As of September 24, 2026 — forty-five days after the August 10 promise and three weeks after the tease's re-affirmation — no Muse Spark 1.2 repository exists on Hugging Face. A direct check of Meta's model organisation on the Hub today returns four repositories, all of them Muse Glimmer variants: there is no Muse Spark checkpoint of any version, no license, no hardware spec and no date. At the September 2 launch, Meta separately indicated it has not decided whether Muse Spark 1.3's weights will follow, and nothing at Connect revised that. The practical situation for a developer deciding whether to plan around a self-hosted Muse Spark is unchanged: the window keeps being stated as "soon," and the only open-weight Muse-family model actually downloadable is Muse Glimmer.

What is confirmed versus what is a guess

Separating the two is the whole job of this piece.

• Confirmed — that the tease exists: Zuckerberg's September 2 Muse Spark 1.3 post ended with the "Next up 🍉 and Muse Spark open weights releases coming soon" line, and @scaling01 relayed it.

• Confirmed — that Muse Spark 1.3 shipped on September 2 at $1.25/$4.25 with a 1M-token context window, and that it is still Meta's newest shipped model after Connect. On Artificial Analysis's Intelligence Index as of September 24, Muse Spark 1.3 (max) scores 48 — behind the closed frontier (Claude Opus 5.5 at 58, Claude Fable 5.1 and GPT-6 Astra at 53) and ahead of the open-weight leaders (Qwen3.8 Max at 45, GLM-5.3 at 45, Kimi K3 at 44). Artificial Analysis has re-based this index more than once since September, so the leaderboard snapshot further down this piece is a September 6 capture on an earlier revision and its figures are not directly comparable to today's.

• Confirmed — that Meta promised Muse Spark 1.2's weights on August 10, 2026, and that as of September 24 the weights have not appeared on Hugging Face; this is checkable by anyone.

• Confirmed — that Muse Glimmer (30B, Apache 2.0) shipped as an open-weight model on August 10, 2026.

• Confirmed — that Meta Connect 2026 took place on September 23–24 and announced no new frontier model; the keynote recaps are public.

• Unverified — that "Watermelon" is the name of the teased model at all; it is a working codename from press reports, and Meta has not confirmed the emoji maps to it.

• Unverified — the October 2026 target, the GPT-5.5-parity claim, the ten-times-compute figure and the reported July pre-training pause; all rest on unnamed sources or press reports.

• Unverified — that Watermelon will be open-weight, or that Muse Spark 1.2 open weights will arrive on any particular date; "soon" is not a date.

The single most likely misinterpretation of the tease — that Meta is about to open-weight its current frontier model — is the one the evidence supports least. Meta has committed to opening the previous checkpoint (Muse Spark 1.2), explicitly left 1.3 undecided, and said nothing about Watermelon's distribution at all.

If both halves land, what actually changes

It is worth being concrete about why this tease matters to someone choosing a model, because the two promises point at different decisions.

An open-weight Muse Spark 1.2 would be the first near-frontier Meta model a team can self-host — no per-token API fee, full data control, weights under whatever license Meta eventually picks. On the September 2026 intelligence board that matters, the top of the Artificial Analysis Intelligence Index is a closed-model crowd — Anthropic's Claude Opus 5.5 at 58, Anthropic's Claude Fable 5.1 and OpenAI's GPT-6 Astra at 53 — with Muse Spark 1.3 (max) at 48, above the current open-weight leaders in the mid-40s. An open-weight Muse Spark 1.2 would be competitive with the top of today's self-hostable field rather than clearly above it — the more consequential scenario is Watermelon. If Watermelon ships open-weight anywhere near the capability Meta has claimed internally, it would sit at or above the top of the open-weight field entirely, and that is precisely why the distribution question is the one to watch, not the benchmark claim.

The other change is subtler and it is already visible: Meta is re-running the Llama playbook in a year where the open-weight frontier has moved to Chinese labs. The open-weight leaders Meta would be chasing — and that an open Muse Spark or Watermelon would have to beat — are not OpenAI or Anthropic; they are the open-weight models from labs like Alibaba, Z.ai and Moonshot that currently define what a team can self-host at the top end, sitting a step below the closed frontier on the same index. Connect added no new evidence that this gap is closing.

That is also the frame for how to act on the tease while it is still a tease. An unshipped model and an undated promise are not something to build a production path around — but the current Muse Spark line is callable today, and a routing layer is the lowest-commitment way to test it against the field. Muse Spark 1.2 is on OrcaRouter at the provider's list price — $1.25 per million input tokens and $4.25 per million output tokens, passed through with 0% markup — so the rate on our side is Meta's own rate, live the same day Meta changes it. Muse Spark 1.3, Muse Glimmer, Muse Voice Transcribe, Muse Image and Muse Video are all not listed on OrcaRouter as of this writing, so this piece does not claim otherwise. Automatic failover is the safe pattern for exactly this situation: point a test path at the newest checkpoint, keep a proven fallback on the same endpoint, and let the router swap in the reliable model if the new one stalls. When — and if — open weights actually land, the honest move is to evaluate them on your own hardware before committing; until then, one API that can reach the whole current board is cheaper than betting on a date Meta has not given.

What to watch next

With the Connect window closed, four events would move this from tease to decision, in rough order of how soon each could happen:

• A Hugging Face repository for Muse Spark 1.2 open weights — the August 10 promise made good; check Meta's model organisations on the Hub rather than waiting for an announcement.

• A license and hardware spec for those weights, which Meta has not yet published; the license determines whether "open weight" means Apache-style freedom or a Llama-style community license.

• An official statement that names Watermelon, confirms it is the 🍉 model, and settles its lineage and distribution — until then the codename is inference, not fact.

• A first independent evaluation of either the reported Watermelon capability or of Muse Spark 1.2's weights once they exist; nothing in the tease is independently verified today.

The reported October window is now the next dated checkpoint, and Meta's Q3 earnings call in late October gives it a natural venue — with the caveat that the same reasoning said Connect would produce something and it did not. Watch the Hub first: open weights are announced with dates far less reliably than they are announced at all, and in this case the promise is forty-five days old with nothing on the registry to show for it.