Mamba-3 Weights Stay Missing After Release

Mamba-3 looks technically important, but the release lands in an uncomfortable middle ground: enough code to study the architecture, not enough weights to reproduce the headline claims cleanly.

• –The paper claims Mamba-3 improves downstream accuracy and inference efficiency at the 1.5B scale, including a stronger MIMO variant with no added decode latency.
• –Together AI says kernels are open-sourced, but that is not the same as releasing the trained checkpoints behind the benchmark tables.
• –The state-spaces GitHub README lists pretrained Hugging Face models for Mamba and Mamba-2, while its Mamba-3 demo uses random tensors for block-level usage.
• –For developers, this makes Mamba-3 more of a research architecture drop than a usable model release until official weights or reproducible training configs appear.