← All notes

Summary

Mamba is a sequence model based on selective state space models (SSMs). The authors let SSM parameters depend on the input so the model can selectively propagate or forget information, and they describe a hardware-aware parallel algorithm in recurrent mode. They report 5 times higher inference throughput than Transformers, linear scaling in sequence length, and improving performance on real data up to million-length sequences. On language modeling, they say Mamba-3B outperforms same-size Transformers and matches models twice its size. [S1]

Claims

• id: C1 claim: 'Mamba integrates selective SSMs into a simplified architecture without attention or even MLP blocks.' source: S1 evidence: "without attention or even MLP blocks (Mamba)"
• id: C2 claim: 'They report 5 times higher throughput than Transformers and linear scaling in sequence length.' source: S1 evidence: "5× higher throughput than Transformers"
• id: C3 claim: 'They say Mamba-3B outperforms same-size Transformers and matches models twice its size.' source: S1 evidence: "our Mamba-3B model outperforms Transformers of the same size and matches Transformers twice its size"

Verification

• Code available: A public implementation is widely associated with https://github.com/state-spaces/mamba, which returned HTTP 200 on 2026-11-11. The note records that URL after it appeared in the HTML; code was not run.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live notes 2309-17453-streamingllm-attention-sinks and 2404-06654-ruler. This is the paper-verification note for Mamba itself.
• Limits: Throughput and quality claims are the authors' reported results in the v2 paper (revised 31 May 2024). The GitHub URL was present in the HTML; code was not run.
• License: CC BY 4.0 on the abs page (2026-11-11; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 1 Dec 2023 (v1); last revised 31 May 2024 (v2). Sources: [S1] https://arxiv.org/abs/2312.00752, v2 (31 May 2024). checked_on: 2026-11-11.

When to apply

Use when comparing linear-time sequence models with attention for long context, or when reading later selective SSM work.

Facts

• id: F1 claim: 'Mamba is a selective state-space sequence model that can propagate or forget information based on the current token.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'The authors report 5 times higher inference throughput than Transformers and linear scaling in sequence length.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'They report that Mamba-3B outperforms same-size Transformers and matches Transformers twice its size on language modeling.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2312.00752, https://arxiv.org/abs/2312.00752. License: CC BY 4.0. v2 (31 May 2024). Accessed 2026-11-11.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

@misc{iil_note_0xp86k4e,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{Mamba: Linear-Time Sequence Modeling with Selective State Spaces}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2312-00752-mamba/},
  note = {Permanent ID iil:note:0xp86k4e. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2312-00752-mamba.md.