Jamba
by AI21 Labs
At a glance
Open model family from AI21 Labs that combines Transformer and state-space (Mamba) architectures.
open-weights language-model mamba
- Licence
- Jamba Open Model License
- Category
- Open Models
- Business model
- Open (open weights/source)
- Availability
- Global
- Launched
- 2024-03
Overview
Jamba is an open model family from AI21 Labs that combines Transformer attention with Mamba state-space layers, enabling very long context windows with efficient inference. Releases include Jamba 1.5 Mini and Large and later revisions, with a 256K-token context. Weights are open under the Jamba Open Model License, and hosted access is sold through AI21's platform and cloud partners.
Compare this tool Suggest a correction
Key features
- Hybrid Transformer-Mamba architecture
- 256K-token context window
- Open weights for Mini and Large variants
- Hosted API via AI21's platform and cloud partners
Use cases
- Long-document analysis and summarization
- Enterprise RAG over large corpora
- Research on hybrid state-space architectures
Pricing
Jamba weights are downloadable under the Jamba Open Model License; managed access through AI21's platform and cloud marketplaces is billed per token.
Record details
Provenance, dates and links are kept here so the main sections stay about the product itself.
Links, dates and provenance
- Record updated
- 2026-09-10
- Model family
- Jamba
- Founded
- 2017
- Headquarters
- IL
- Canonical URL
- https://globalaiproductindex.com/products/jamba/
Selection profile
Every field says where its value comes from. "Unknown" means we have not confirmed it — it does not mean the product lacks the feature.
Selection profile11 fields2 unknown
- Tasks it fits
- Running models yourself
Derived from the record: category "open-models".
Suggest a correction - How you use it
- Model weights
Derived from record tag matching /^open-weights$/.
Suggest a correction - Runs on
- Windows
Derived from platform words in the record description/features/tags.
Suggest a correction - Where it runs
- Local machine
Open-weight models are run by the user, not served by the vendor.
Suggest a correction - Works offline
- Partly
Local execution implies the model can run without a network once downloaded, but the vendor does not state this in the record.
Suggest a correction - Official service region
- Global
Mirrors the availability field already on the record ("global").
Suggest a correction - Public API
- Yes
The record mentions an API surface.
Suggest a correction - Licence
- Jamba Open Model License
Mirrors the licence already recorded on the product record; not re-verified against the licence text by this migration.
Suggest a correction - Commercial use
- Not recorded
No licence terms recorded, so commercial use cannot be asserted.
Suggest a correction - Interface languages
- Not recorded
No vendor page read in this pass states the product's own interface languages; the key exists so the field can be answered later rather than left absent.
Suggest a correction - Billing
- Open source Free Free allowance: Yes Subscription required: No
Prices are the vendor’s advertised entry price and exclude taxes; always confirm on the vendor’s own page.
Suggest a correction
Where each value comes from
- Tasks it fits
- Our inference
- How you use it
- Our inference
- Runs on
- Our inference
- Where it runs
- Our inference
- Works offline
- Our inference
- Official service region
- Our inference
- Public API
- Our inference
- Licence
- Our inference
- Commercial use
- Our inference
- Interface languages
- Unknown
Frequently asked questions
What is Jamba?
Jamba is AI21's open model family mixing Transformer and Mamba layers for efficient long-context language modeling.
How long is Jamba's context window?
Jamba models support a 256K-token context window.
Are Jamba models open?
Yes, weights are released under the Jamba Open Model License, with commercial hosted options from AI21.
What does Jamba not record?
No Jamba documentation or pricing page is recorded, because no same-site page confirming either could be identified.
Similar products
- Nemotron — Family of open language models released by NVIDIA for enterprise and research use.
- Command R — Cohere's open-weight LLM family optimized for retrieval-augmented generation and tool use; succeeded by Command A as Cohere's flagship.
- DBRX — Open-weight mixture-of-experts LLM released by Databricks in 2024; retired from Databricks serving in 2025, though the weights remain available.
- Falcon — Family of open-weight large language models developed by the Technology Innovation Institute (TII).
- Granite — IBM's family of open-weight language models aimed at enterprise and code tasks.
- Baichuan — Open-weight bilingual LLM family from Beijing-based Baichuan Intelligence, now focused on medical AI with its Baichuan-M reasoning models.
Related guides
Sources
This record was last reviewed on 2026-09-10.
Machine-readable record: /api/products/jamba.json · Spot an error? Suggest a correction