Mamba

by CMU / Princeton (Gu, Dao)

System Card

OrganizationCMU / Princeton (Gu, Dao)

Released2023-12

Architectureexternal-memory-network / Selective state-space model (input-dependent SSM)

DetailsMakes SSM parameters input-dependent, allowing selective information propagation. Replaces attention/MLP blocks with a unified selective SSM block for linear-time sequence modeling and constant-size recurrent state.

Parameters—

Domainlong-context

Open SourceYes

PaperView Paper

CodeRepository

colm-2024ssmlinear-timeselectiverecurrent

Capability Profile

Benchmark Scores

6 of 14 benchmarks

Long-Context Retrieval

5/5

77.890p

75.138p

81.191p

603p

8588p

Multi-Turn Recall

0/2

LoCoMo

no data

MemoryBank

no data

Cross-Session Memory

0/1

LongMemEval

no data

Multi-Hop QA

1/3

BABILong

82.597p

MultiHop-RAG

no data

HotpotQA

no data

Agent Task Memory

0/1

AgentBench-Mem

no data

Personalization

0/1

PerLTQA

no data

Factuality / Grounding

0/1

RAGAS

no data

Sources:Mamba paper (arXiv:2312.00752); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)Mamba paper (arXiv:2312.00752); evaluated on InfiniteBench: Extending Long Context Evaluation Beyond 100K Tokens (Tsinghua / OpenBMB, 2402)Mamba paper (arXiv:2312.00752); evaluated on LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding (Tsinghua KEG, 2308)Mamba paper (arXiv:2312.00752); evaluated on LooGLE: Can Long-Context Language Models Understand Long Contexts? (Peking University, 2311)Mamba paper (arXiv:2312.00752); evaluated on Needle in a Haystack (Greg Kamradt, 2024)Mamba paper (arXiv:2312.00752); evaluated on RULER: What's the Real Context Size of Your Long-Context Language Models (NVIDIA, 2404)