Bamba 9B
IBM's open hybrid Mamba2-transformer model: 9.8B parameters with a fraction of the KV-cache of a dense transformer.
About
Bamba-9B is a hybrid language model built by IBM Research with Princeton, CMU, and UIUC, combining Mamba2 state-space layers with a small number of shared transformer attention blocks. The first version trained on 2.2 trillion tokens over Christmas 2024; the v2 release in April 2025 added another trillion tokens plus an annealing phase. Everything around it is open: the Apache-2.0 weights, the training recipes, the data loader, and support in transformers, vLLM, TRL, and llama.cpp.
The design target is the transformer's quadratic KV-cache bottleneck: IBM reports inference roughly 2-2.5x faster than similar-size transformers in vLLM. On the provider's own benchmark table the v1 base model scores 60.77 on MMLU (5-shot), 63.23 on ARC-C, and 81.8 on HellaSwag, and the team reports v2 beating Llama 3.1 8B on OpenLLM leaderboard averages despite far less training data. Context length is a modest 4,096 tokens per the model card, which limits its use for long-document work.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- MMLU (5-shot): 60.77, ARC-C (25-shot): 63.23, GSM8K (5-shot): 36.77, HellaSwag (10-shot): 81.8 (provider-reported, v1 base model card)
- v2 reported to outperform Llama 3.1 8B on OpenLLM v1/v2 leaderboard averages (provider-reported, foundation-model-stack blog)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: IBM Research release with full training lineage (2.2T tokens v1, +1T v2), Apache-2.0 weights, open datasets and training code; reproducible eval scripts published. Open weights. Release facts corroborated across IBM Research, the foundation-model-stack blog, and press. Strengths: genuine architectural contribution, strong reproducibility story. Docked for: (1) benchmark comparisons are provider-run with no independent replication found; (2) 4,096-token context is small for 2025-era workloads and the model is base-only (SFT/DPO marked coming soon).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · IBM Research + Princeton/CMU/UIUC; v1 Dec 2024 (2.2T tokens), v2 Apr 2025 (+1T tokens)
- · Apache-2.0 weights; open training recipes, data loader, published eval scripts
- · Provider benchmark table: MMLU 60.77, ARC-C 63.23, HellaSwag 81.8 (v1 base)
- · 4,096-token context is small; SFT/DPO variants marked coming soon