Mamba 2.8B

Mamba 2.8B is a language model built around a selective state-space architecture rather than a Transformer. A state-space model processes a sequence through a compact evolving internal state, offering a different route to handling long sequences than Transformer’s all-to-all attention. Mamba was introduced by Albert Gu and Tri Dao; the architectural distinction matters more here than a corporate brand. It still performs familiar language-model work, predicting and generating text, so ‘not a Transformer’ does not mean ‘not an LLM.’ TT-Metal lists the 2.8B model on a single n150 Wormhole device at batch 32. The matrix also notes that Mamba’s reported setup can accept any sequence length, unlike the fixed measurement condition stated for its Transformer LLM rows.