Fractale-350M-base model released with trained fast-weight memory

The 350M parameter model introduces a trained fast-weight memory as an alternative to long context, developed by a solo researcher on a single RTX 3090. The release includes the paper and full research log on GitHub.