Key Info
A reviewer who had early access to Meta's new model codenamed "Watermelon" shared impressions, noting it resisted trick questions, followed instructions precisely, and functioned as a coding agent that wrote, ran, and checked its own code.
Highlights
- Reviewer reports the model handled trick questions without being fooled
- It followed instructions to the word
- As a coding agent it wrote, ran, and checked its own code
- Benchmark results matched the reviewer's hands-on experience
- Creative writing was described as still a bit dry
- Reviewer calls the overall result "very very very impressive"