xAI shipped Grok 4.7 on September 21, 2026. Google's Gemini 3.8 Flash shipped August 26 and targets consumer speed, not coding depth (xAI; Google, September 2026).
How to follow the keynote
The keynote is livestreamed at meta.com/connect and via Meta's developer YouTube channel. Meta will publish full session recordings and the developer API updates within 24 hours of the keynote.
The Muse Spark 1.2 open-weights release and the Watermelon model family announcement are the two highest-volume keywords to watch on the day (Meta, September 2026).
What the Watermelon family announcement could include
If Meta follows the pattern set by Muse Spark 1.3 in September 2026, expect three things from a Watermelon reveal.
First, the headline model announcement. A single frontier-class API model - probably a 70B+ dense or sparse mixture-of-experts architecture, tuned for coding and knowledge work. Pricing will likely mirror or undercut the Muse Spark 1.3 standard tier ($1.25/$4.25 per million).
Second, a multimodal expansion. Watermelon is positioned as a step beyond Muse Spark's text-and-image coverage. Expect native audio understanding (rivaling Google's Gemini audio), video input, and possibly native video generation via Muse Video.
Third, a developer-platform reveal. The Video Playground and Voice Playground sections on the developer portal are placeholders for Watermelon-era capabilities. Pricing for the Playground endpoints and the Connectors revenue share will likely be announced alongside the model.
If the keynote is a soft-launch, expect Watermelon weights behind a gated program rather than open release. Meta has been more cautious about open weights since the Muse 1.2 safety fine-tuning push slipped from August into September.
Why this matters for frontier competition
Meta's September 2026 push signals the company's confidence in the frontier race. Wang's August framing of Watermelon as extremely competitive, paired with the open-weights roadmap, is a direct response to OpenAI's closed-API strategy and Anthropic's pacing-the-frontier posture.
For enterprise buyers, the next four weeks resolve most of the open questions. Pricing, open weights availability, multimodal coverage, and tool-call benchmarks will all be updated across the major labs.
For self-hosted teams, Muse Spark 1.2 open weights are the single highest-impact event. Watch Meta's GitHub and Hugging Face accounts for the announcement.
Muse Glimmer 30B is already available as open weights - it is a 30B dense distillation that runs on consumer hardware. Teams evaluating the family should benchmark Muse Glimmer 30B first as a free reference, then test Muse Spark 1.3 API access.
The Contributor tier ($0.10 input / $0.20 output per million tokens, with training-data opt-in) remains the cheapest frontier-class API in production - and is a unique moat against OpenAI and Anthropic for teams comfortable sharing training data.
Internal links
For the Muse Spark 1.3 release details, see our Muse Spark 1.3 launch post.
For the IPO and capex backdrop driving the frontier race, see our OpenAI Anthropic IPO preparations post.
For a three-way comparison with Claude and Gemini, see our ChatGPT vs Claude vs Gemini guide.