Jamba 1.5 Large
AI21 LabsAI21-Jamba-1.5-Large (ai21labs/AI21-Jamba-1.5-Large on HF), hybrid SSM-Transformer (Mamba+Attention) MoE, 398B total / 94B active, 256K context, instruction-tuned chat model with tool use / grounded-RAG / JSON mode. Announced 22 Aug 2024 under the Jamba Open Model License. Stale: superseded by Jamba 1.6 (Mar 2025) and the Jamba2/Jamba Reasoning 3B generation now headlined on ai21.com/jamba. Recategorized from base_pretrained -> finetuned_chat: this is the instruct/chat SKU (post-trained for downstream chat/agentic use), not a raw base model.
Recategorized from base_pretrained -> finetuned_chat; this is the instruct/chat SKU, not a raw base model. The repo is gated `auto`, so openness was read from the model card and the license body rather than a download. Verified 2026-08-13 via the HF model card, the Jamba Open Model License body and the AI21 release blog.
Openness
3 high confidence- weights
- open(downloadable on HF, contact-info gate)
- data
- closed
- code
- partial(inference/usage only, no training pipeline)
- license
- Jamba-Open-Model-License(non-OSI: $50M annual-revenue commercial cap + AUP + naming/attribution requirements)
Weights are downloadable under the Jamba Open Model License, which is not OSI-approved and carries a $50M annual-revenue cap that bars large-entity commercial use; the "Open Model" branding overstates it. A license that bounds commercial use rather than forbidding it caps at 3 rather than 2: the test at that boundary is whether commercial use is permitted at all, and this one permits it for everyone inside the bound, where a license that forbids commerce outright - CC-BY-NC, Mistral's Non-Production License - stays at 2. Nine other products carry a bounded license and all score 3.
- https://huggingface.co/ai21labs/AI21-Jamba-1.5-Large recorded 2026-08-13
"The models are released under the Jamba Open Model License, a permissive license allowing full research use and commercial use under the license terms"; `license:other` in the repo tags with `"gated":"auto"`; downloadable safetensors; inference and usage instructions only, no training corpus or pipeline
- https://www.ai21.com/licenses/jamba-open-model-license recorded 2026-08-13
license text - "if at any time You and/or Your affiliate(s), whether individually or in aggregate, generate more than fifty million US Dollars (USD $50,000,000) in annual revenue ... then You" must license separately; non-OSI
Adoption
1 high confidenceThe scored release, ai21labs/AI21-Jamba-1.5-Large, reads 45 Hugging Face downloads in the trailing 30 days, which bands at level 1 (<10K). That repository is declared on this product, and adoption is summed across the releases a tier covers.
- https://huggingface.co/api/models/ai21labs/AI21-Jamba-1.5-Large recorded 2026-08-12
45 downloads in the trailing 30 days for ai21labs/AI21-Jamba-1.5-Large
Capability
3 medium confidenceA 2024 hybrid SSM-Transformer that led open models on long context at release and is now mid-tier against the 2026 frontier. The HF card's benchmark table gives Jamba 1.5 Large Arena Hard 65.4, MMLU (CoT) 81.2 and MMLU Pro (CoT) 53.5, and AI21's launch post frames the 65.4 as outpacing both Llama 3.1 70B and 405B - a strong placement for its generation, and well short of the 2026 reasoning and coding frontier this category scores against.
- https://www.ai21.com/blog/announcing-jamba-model-family/ recorded 2026-08-13
"Jamba 1.5 Large, with a score of 65.4, outpaces both Llama 3.1 70B and 405B" on Arena Hard; Jamba Open Model License
- https://huggingface.co/ai21labs/AI21-Jamba-1.5-Large recorded 2026-08-13
benchmark table, Jamba 1.5 Large column - Arena Hard 65.4, Wild Bench 48.5, MMLU (CoT) 81.2, MMLU Pro (CoT) 53.5; a RULER Benchmark effective-context-length section
Verified 2026-08-12