Flash Onyx 2.2 Is Out, and It Finally Finishes the Job

Flash Onyx 2.2, the model behind FLASH CLI, has been released on Ollama. The upgrade from 2.1 aimed to improve the model's accuracy and functionality. There are two sizes available, both running on regular consumer hardware – 12b and the flagship 31b, which requires a real GPU. To use Flash Onyx 2.2, pull it with the command "ollama pull Natuworkguy/flash-onyx-2.2:12b" and "ollama pull Natuworkguy/flash-onyx-2.2:31b". Set the environment variable "MODEL=flash-onyx-2.2:12b".

Several key changes were introduced in 2.2. First, the model now reads the request and satisfies each part, providing answers at the appropriate level. For example, "Running is the floor, not the finish" gets a yes, while "What actually changed" gets a detailed explanation. It also stops confusing questions with instructions and offers appropriate responses, such as splitting a file or choosing a stack.

2.2 has improved the Manim library by building with VGroup and .arrange() instead of hand-coded .move_to() coordinates, computing geometry from real points, and assigning deliberate run_time and rate_func to beats. This prevents hallucinations and ensures a better user experience. The model also bans certain file names like "untitled", "output", "script", and "v2" to maintain consistency.

Another significant improvement is in handling contracts and documents. 2.2 reads the entire document, reviews it from different sides, and ranks findings accordingly. It also identifies missing elements and provides clear recommendations, making it more useful for users. The model now generates more comprehensive and accurate responses across various use cases, making it a more reliable tool for developers and users alike.