Early Access Leak: Meta's "Watermelon" Follows Instructions Precisely, Self-Checks Code
petrusenko_max · x · 2026-10-13
Citing kimmonismus (unverified): he has had early access to Meta's new "Watermelon" model. Initial testing shows trick questions don't fool it, it follows instructions to the word, and as a coding agent it writes, runs and verifies its own code. Creative writing reads a bit dry, but performance matches the benchmarks — which he hadn't seen in advance and found genuinely surprising.
Related event: Blogger Leaks Early Test of Meta's Watermelon Model(3 posts)→
More from Models
- 29 bugs in benchmark history have never been fixed by any model, says Pawel Huryn — PawelHuryn · 2026-10-13
- Dan Shipper kicks off full runs of Wonder on AI&) senior engineer benchmark — danshipper · 2026-10-13
- Codex users hit 'model at capacity' on Sol 6.1 and scramble for fallbacks — snug-crackle-policy · 2026-10-13
- Real-world test: Muse Wonder 1.0 fixes 31.7/105 bugs, 50% pricier than Spark 1.3 — PawelHuryn · 2026-10-13
- Claude's ban wave hits Japanese users, sparking online mourning — lxfater · 2026-10-13
- Choosing Models by Public Benchmarks and Vibes Is 'ngmi', Argues Dev for Internal Evals — A_K_Nain · 2026-10-13